cachepc-qemu

Fork of AMDESE/qemu with changes for cachepc side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-qemu
Log | Files | Refs | Submodules | LICENSE | sfeed.txt

mve_helper.c (166652B)


      1/*
      2 * M-profile MVE Operations
      3 *
      4 * Copyright (c) 2021 Linaro, Ltd.
      5 *
      6 * This library is free software; you can redistribute it and/or
      7 * modify it under the terms of the GNU Lesser General Public
      8 * License as published by the Free Software Foundation; either
      9 * version 2.1 of the License, or (at your option) any later version.
     10 *
     11 * This library is distributed in the hope that it will be useful,
     12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
     13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
     14 * Lesser General Public License for more details.
     15 *
     16 * You should have received a copy of the GNU Lesser General Public
     17 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
     18 */
     19
     20#include "qemu/osdep.h"
     21#include "cpu.h"
     22#include "internals.h"
     23#include "vec_internal.h"
     24#include "exec/helper-proto.h"
     25#include "exec/cpu_ldst.h"
     26#include "exec/exec-all.h"
     27#include "tcg/tcg.h"
     28#include "fpu/softfloat.h"
     29
     30static uint16_t mve_eci_mask(CPUARMState *env)
     31{
     32    /*
     33     * Return the mask of which elements in the MVE vector correspond
     34     * to beats being executed. The mask has 1 bits for executed lanes
     35     * and 0 bits where ECI says this beat was already executed.
     36     */
     37    int eci;
     38
     39    if ((env->condexec_bits & 0xf) != 0) {
     40        return 0xffff;
     41    }
     42
     43    eci = env->condexec_bits >> 4;
     44    switch (eci) {
     45    case ECI_NONE:
     46        return 0xffff;
     47    case ECI_A0:
     48        return 0xfff0;
     49    case ECI_A0A1:
     50        return 0xff00;
     51    case ECI_A0A1A2:
     52    case ECI_A0A1A2B0:
     53        return 0xf000;
     54    default:
     55        g_assert_not_reached();
     56    }
     57}
     58
     59static uint16_t mve_element_mask(CPUARMState *env)
     60{
     61    /*
     62     * Return the mask of which elements in the MVE vector should be
     63     * updated. This is a combination of multiple things:
     64     *  (1) by default, we update every lane in the vector
     65     *  (2) VPT predication stores its state in the VPR register;
     66     *  (3) low-overhead-branch tail predication will mask out part
     67     *      the vector on the final iteration of the loop
     68     *  (4) if EPSR.ECI is set then we must execute only some beats
     69     *      of the insn
     70     * We combine all these into a 16-bit result with the same semantics
     71     * as VPR.P0: 0 to mask the lane, 1 if it is active.
     72     * 8-bit vector ops will look at all bits of the result;
     73     * 16-bit ops will look at bits 0, 2, 4, ...;
     74     * 32-bit ops will look at bits 0, 4, 8 and 12.
     75     * Compare pseudocode GetCurInstrBeat(), though that only returns
     76     * the 4-bit slice of the mask corresponding to a single beat.
     77     */
     78    uint16_t mask = FIELD_EX32(env->v7m.vpr, V7M_VPR, P0);
     79
     80    if (!(env->v7m.vpr & R_V7M_VPR_MASK01_MASK)) {
     81        mask |= 0xff;
     82    }
     83    if (!(env->v7m.vpr & R_V7M_VPR_MASK23_MASK)) {
     84        mask |= 0xff00;
     85    }
     86
     87    if (env->v7m.ltpsize < 4 &&
     88        env->regs[14] <= (1 << (4 - env->v7m.ltpsize))) {
     89        /*
     90         * Tail predication active, and this is the last loop iteration.
     91         * The element size is (1 << ltpsize), and we only want to process
     92         * loopcount elements, so we want to retain the least significant
     93         * (loopcount * esize) predicate bits and zero out bits above that.
     94         */
     95        int masklen = env->regs[14] << env->v7m.ltpsize;
     96        assert(masklen <= 16);
     97        uint16_t ltpmask = masklen ? MAKE_64BIT_MASK(0, masklen) : 0;
     98        mask &= ltpmask;
     99    }
    100
    101    /*
    102     * ECI bits indicate which beats are already executed;
    103     * we handle this by effectively predicating them out.
    104     */
    105    mask &= mve_eci_mask(env);
    106    return mask;
    107}
    108
    109static void mve_advance_vpt(CPUARMState *env)
    110{
    111    /* Advance the VPT and ECI state if necessary */
    112    uint32_t vpr = env->v7m.vpr;
    113    unsigned mask01, mask23;
    114    uint16_t inv_mask;
    115    uint16_t eci_mask = mve_eci_mask(env);
    116
    117    if ((env->condexec_bits & 0xf) == 0) {
    118        env->condexec_bits = (env->condexec_bits == (ECI_A0A1A2B0 << 4)) ?
    119            (ECI_A0 << 4) : (ECI_NONE << 4);
    120    }
    121
    122    if (!(vpr & (R_V7M_VPR_MASK01_MASK | R_V7M_VPR_MASK23_MASK))) {
    123        /* VPT not enabled, nothing to do */
    124        return;
    125    }
    126
    127    /* Invert P0 bits if needed, but only for beats we actually executed */
    128    mask01 = FIELD_EX32(vpr, V7M_VPR, MASK01);
    129    mask23 = FIELD_EX32(vpr, V7M_VPR, MASK23);
    130    /* Start by assuming we invert all bits corresponding to executed beats */
    131    inv_mask = eci_mask;
    132    if (mask01 <= 8) {
    133        /* MASK01 says don't invert low half of P0 */
    134        inv_mask &= ~0xff;
    135    }
    136    if (mask23 <= 8) {
    137        /* MASK23 says don't invert high half of P0 */
    138        inv_mask &= ~0xff00;
    139    }
    140    vpr ^= inv_mask;
    141    /* Only update MASK01 if beat 1 executed */
    142    if (eci_mask & 0xf0) {
    143        vpr = FIELD_DP32(vpr, V7M_VPR, MASK01, mask01 << 1);
    144    }
    145    /* Beat 3 always executes, so update MASK23 */
    146    vpr = FIELD_DP32(vpr, V7M_VPR, MASK23, mask23 << 1);
    147    env->v7m.vpr = vpr;
    148}
    149
    150/* For loads, predicated lanes are zeroed instead of keeping their old values */
    151#define DO_VLDR(OP, MSIZE, LDTYPE, ESIZE, TYPE)                         \
    152    void HELPER(mve_##OP)(CPUARMState *env, void *vd, uint32_t addr)    \
    153    {                                                                   \
    154        TYPE *d = vd;                                                   \
    155        uint16_t mask = mve_element_mask(env);                          \
    156        uint16_t eci_mask = mve_eci_mask(env);                          \
    157        unsigned b, e;                                                  \
    158        /*                                                              \
    159         * R_SXTM allows the dest reg to become UNKNOWN for abandoned   \
    160         * beats so we don't care if we update part of the dest and     \
    161         * then take an exception.                                      \
    162         */                                                             \
    163        for (b = 0, e = 0; b < 16; b += ESIZE, e++) {                   \
    164            if (eci_mask & (1 << b)) {                                  \
    165                d[H##ESIZE(e)] = (mask & (1 << b)) ?                    \
    166                    cpu_##LDTYPE##_data_ra(env, addr, GETPC()) : 0;     \
    167            }                                                           \
    168            addr += MSIZE;                                              \
    169        }                                                               \
    170        mve_advance_vpt(env);                                           \
    171    }
    172
    173#define DO_VSTR(OP, MSIZE, STTYPE, ESIZE, TYPE)                         \
    174    void HELPER(mve_##OP)(CPUARMState *env, void *vd, uint32_t addr)    \
    175    {                                                                   \
    176        TYPE *d = vd;                                                   \
    177        uint16_t mask = mve_element_mask(env);                          \
    178        unsigned b, e;                                                  \
    179        for (b = 0, e = 0; b < 16; b += ESIZE, e++) {                   \
    180            if (mask & (1 << b)) {                                      \
    181                cpu_##STTYPE##_data_ra(env, addr, d[H##ESIZE(e)], GETPC()); \
    182            }                                                           \
    183            addr += MSIZE;                                              \
    184        }                                                               \
    185        mve_advance_vpt(env);                                           \
    186    }
    187
    188DO_VLDR(vldrb, 1, ldub, 1, uint8_t)
    189DO_VLDR(vldrh, 2, lduw, 2, uint16_t)
    190DO_VLDR(vldrw, 4, ldl, 4, uint32_t)
    191
    192DO_VSTR(vstrb, 1, stb, 1, uint8_t)
    193DO_VSTR(vstrh, 2, stw, 2, uint16_t)
    194DO_VSTR(vstrw, 4, stl, 4, uint32_t)
    195
    196DO_VLDR(vldrb_sh, 1, ldsb, 2, int16_t)
    197DO_VLDR(vldrb_sw, 1, ldsb, 4, int32_t)
    198DO_VLDR(vldrb_uh, 1, ldub, 2, uint16_t)
    199DO_VLDR(vldrb_uw, 1, ldub, 4, uint32_t)
    200DO_VLDR(vldrh_sw, 2, ldsw, 4, int32_t)
    201DO_VLDR(vldrh_uw, 2, lduw, 4, uint32_t)
    202
    203DO_VSTR(vstrb_h, 1, stb, 2, int16_t)
    204DO_VSTR(vstrb_w, 1, stb, 4, int32_t)
    205DO_VSTR(vstrh_w, 2, stw, 4, int32_t)
    206
    207#undef DO_VLDR
    208#undef DO_VSTR
    209
    210/*
    211 * Gather loads/scatter stores. Here each element of Qm specifies
    212 * an offset to use from the base register Rm. In the _os_ versions
    213 * that offset is scaled by the element size.
    214 * For loads, predicated lanes are zeroed instead of retaining
    215 * their previous values.
    216 */
    217#define DO_VLDR_SG(OP, LDTYPE, ESIZE, TYPE, OFFTYPE, ADDRFN, WB)        \
    218    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm,         \
    219                          uint32_t base)                                \
    220    {                                                                   \
    221        TYPE *d = vd;                                                   \
    222        OFFTYPE *m = vm;                                                \
    223        uint16_t mask = mve_element_mask(env);                          \
    224        uint16_t eci_mask = mve_eci_mask(env);                          \
    225        unsigned e;                                                     \
    226        uint32_t addr;                                                  \
    227        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE, eci_mask >>= ESIZE) { \
    228            if (!(eci_mask & 1)) {                                      \
    229                continue;                                               \
    230            }                                                           \
    231            addr = ADDRFN(base, m[H##ESIZE(e)]);                        \
    232            d[H##ESIZE(e)] = (mask & 1) ?                               \
    233                cpu_##LDTYPE##_data_ra(env, addr, GETPC()) : 0;         \
    234            if (WB) {                                                   \
    235                m[H##ESIZE(e)] = addr;                                  \
    236            }                                                           \
    237        }                                                               \
    238        mve_advance_vpt(env);                                           \
    239    }
    240
    241/* We know here TYPE is unsigned so always the same as the offset type */
    242#define DO_VSTR_SG(OP, STTYPE, ESIZE, TYPE, ADDRFN, WB)                 \
    243    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm,         \
    244                          uint32_t base)                                \
    245    {                                                                   \
    246        TYPE *d = vd;                                                   \
    247        TYPE *m = vm;                                                   \
    248        uint16_t mask = mve_element_mask(env);                          \
    249        uint16_t eci_mask = mve_eci_mask(env);                          \
    250        unsigned e;                                                     \
    251        uint32_t addr;                                                  \
    252        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE, eci_mask >>= ESIZE) { \
    253            if (!(eci_mask & 1)) {                                      \
    254                continue;                                               \
    255            }                                                           \
    256            addr = ADDRFN(base, m[H##ESIZE(e)]);                        \
    257            if (mask & 1) {                                             \
    258                cpu_##STTYPE##_data_ra(env, addr, d[H##ESIZE(e)], GETPC()); \
    259            }                                                           \
    260            if (WB) {                                                   \
    261                m[H##ESIZE(e)] = addr;                                  \
    262            }                                                           \
    263        }                                                               \
    264        mve_advance_vpt(env);                                           \
    265    }
    266
    267/*
    268 * 64-bit accesses are slightly different: they are done as two 32-bit
    269 * accesses, controlled by the predicate mask for the relevant beat,
    270 * and with a single 32-bit offset in the first of the two Qm elements.
    271 * Note that for QEMU our IMPDEF AIRCR.ENDIANNESS is always 0 (little).
    272 * Address writeback happens on the odd beats and updates the address
    273 * stored in the even-beat element.
    274 */
    275#define DO_VLDR64_SG(OP, ADDRFN, WB)                                    \
    276    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm,         \
    277                          uint32_t base)                                \
    278    {                                                                   \
    279        uint32_t *d = vd;                                               \
    280        uint32_t *m = vm;                                               \
    281        uint16_t mask = mve_element_mask(env);                          \
    282        uint16_t eci_mask = mve_eci_mask(env);                          \
    283        unsigned e;                                                     \
    284        uint32_t addr;                                                  \
    285        for (e = 0; e < 16 / 4; e++, mask >>= 4, eci_mask >>= 4) {      \
    286            if (!(eci_mask & 1)) {                                      \
    287                continue;                                               \
    288            }                                                           \
    289            addr = ADDRFN(base, m[H4(e & ~1)]);                         \
    290            addr += 4 * (e & 1);                                        \
    291            d[H4(e)] = (mask & 1) ? cpu_ldl_data_ra(env, addr, GETPC()) : 0; \
    292            if (WB && (e & 1)) {                                        \
    293                m[H4(e & ~1)] = addr - 4;                               \
    294            }                                                           \
    295        }                                                               \
    296        mve_advance_vpt(env);                                           \
    297    }
    298
    299#define DO_VSTR64_SG(OP, ADDRFN, WB)                                    \
    300    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm,         \
    301                          uint32_t base)                                \
    302    {                                                                   \
    303        uint32_t *d = vd;                                               \
    304        uint32_t *m = vm;                                               \
    305        uint16_t mask = mve_element_mask(env);                          \
    306        uint16_t eci_mask = mve_eci_mask(env);                          \
    307        unsigned e;                                                     \
    308        uint32_t addr;                                                  \
    309        for (e = 0; e < 16 / 4; e++, mask >>= 4, eci_mask >>= 4) {      \
    310            if (!(eci_mask & 1)) {                                      \
    311                continue;                                               \
    312            }                                                           \
    313            addr = ADDRFN(base, m[H4(e & ~1)]);                         \
    314            addr += 4 * (e & 1);                                        \
    315            if (mask & 1) {                                             \
    316                cpu_stl_data_ra(env, addr, d[H4(e)], GETPC());          \
    317            }                                                           \
    318            if (WB && (e & 1)) {                                        \
    319                m[H4(e & ~1)] = addr - 4;                               \
    320            }                                                           \
    321        }                                                               \
    322        mve_advance_vpt(env);                                           \
    323    }
    324
    325#define ADDR_ADD(BASE, OFFSET) ((BASE) + (OFFSET))
    326#define ADDR_ADD_OSH(BASE, OFFSET) ((BASE) + ((OFFSET) << 1))
    327#define ADDR_ADD_OSW(BASE, OFFSET) ((BASE) + ((OFFSET) << 2))
    328#define ADDR_ADD_OSD(BASE, OFFSET) ((BASE) + ((OFFSET) << 3))
    329
    330DO_VLDR_SG(vldrb_sg_sh, ldsb, 2, int16_t, uint16_t, ADDR_ADD, false)
    331DO_VLDR_SG(vldrb_sg_sw, ldsb, 4, int32_t, uint32_t, ADDR_ADD, false)
    332DO_VLDR_SG(vldrh_sg_sw, ldsw, 4, int32_t, uint32_t, ADDR_ADD, false)
    333
    334DO_VLDR_SG(vldrb_sg_ub, ldub, 1, uint8_t, uint8_t, ADDR_ADD, false)
    335DO_VLDR_SG(vldrb_sg_uh, ldub, 2, uint16_t, uint16_t, ADDR_ADD, false)
    336DO_VLDR_SG(vldrb_sg_uw, ldub, 4, uint32_t, uint32_t, ADDR_ADD, false)
    337DO_VLDR_SG(vldrh_sg_uh, lduw, 2, uint16_t, uint16_t, ADDR_ADD, false)
    338DO_VLDR_SG(vldrh_sg_uw, lduw, 4, uint32_t, uint32_t, ADDR_ADD, false)
    339DO_VLDR_SG(vldrw_sg_uw, ldl, 4, uint32_t, uint32_t, ADDR_ADD, false)
    340DO_VLDR64_SG(vldrd_sg_ud, ADDR_ADD, false)
    341
    342DO_VLDR_SG(vldrh_sg_os_sw, ldsw, 4, int32_t, uint32_t, ADDR_ADD_OSH, false)
    343DO_VLDR_SG(vldrh_sg_os_uh, lduw, 2, uint16_t, uint16_t, ADDR_ADD_OSH, false)
    344DO_VLDR_SG(vldrh_sg_os_uw, lduw, 4, uint32_t, uint32_t, ADDR_ADD_OSH, false)
    345DO_VLDR_SG(vldrw_sg_os_uw, ldl, 4, uint32_t, uint32_t, ADDR_ADD_OSW, false)
    346DO_VLDR64_SG(vldrd_sg_os_ud, ADDR_ADD_OSD, false)
    347
    348DO_VSTR_SG(vstrb_sg_ub, stb, 1, uint8_t, ADDR_ADD, false)
    349DO_VSTR_SG(vstrb_sg_uh, stb, 2, uint16_t, ADDR_ADD, false)
    350DO_VSTR_SG(vstrb_sg_uw, stb, 4, uint32_t, ADDR_ADD, false)
    351DO_VSTR_SG(vstrh_sg_uh, stw, 2, uint16_t, ADDR_ADD, false)
    352DO_VSTR_SG(vstrh_sg_uw, stw, 4, uint32_t, ADDR_ADD, false)
    353DO_VSTR_SG(vstrw_sg_uw, stl, 4, uint32_t, ADDR_ADD, false)
    354DO_VSTR64_SG(vstrd_sg_ud, ADDR_ADD, false)
    355
    356DO_VSTR_SG(vstrh_sg_os_uh, stw, 2, uint16_t, ADDR_ADD_OSH, false)
    357DO_VSTR_SG(vstrh_sg_os_uw, stw, 4, uint32_t, ADDR_ADD_OSH, false)
    358DO_VSTR_SG(vstrw_sg_os_uw, stl, 4, uint32_t, ADDR_ADD_OSW, false)
    359DO_VSTR64_SG(vstrd_sg_os_ud, ADDR_ADD_OSD, false)
    360
    361DO_VLDR_SG(vldrw_sg_wb_uw, ldl, 4, uint32_t, uint32_t, ADDR_ADD, true)
    362DO_VLDR64_SG(vldrd_sg_wb_ud, ADDR_ADD, true)
    363DO_VSTR_SG(vstrw_sg_wb_uw, stl, 4, uint32_t, ADDR_ADD, true)
    364DO_VSTR64_SG(vstrd_sg_wb_ud, ADDR_ADD, true)
    365
    366/*
    367 * Deinterleaving loads/interleaving stores.
    368 *
    369 * For these helpers we are passed the index of the first Qreg
    370 * (VLD2/VST2 will also access Qn+1, VLD4/VST4 access Qn .. Qn+3)
    371 * and the value of the base address register Rn.
    372 * The helpers are specialized for pattern and element size, so
    373 * for instance vld42h is VLD4 with pattern 2, element size MO_16.
    374 *
    375 * These insns are beatwise but not predicated, so we must honour ECI,
    376 * but need not look at mve_element_mask().
    377 *
    378 * The pseudocode implements these insns with multiple memory accesses
    379 * of the element size, but rules R_VVVG and R_FXDM permit us to make
    380 * one 32-bit memory access per beat.
    381 */
    382#define DO_VLD4B(OP, O1, O2, O3, O4)                                    \
    383    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    384                          uint32_t base)                                \
    385    {                                                                   \
    386        int beat, e;                                                    \
    387        uint16_t mask = mve_eci_mask(env);                              \
    388        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    389        uint32_t addr, data;                                            \
    390        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    391            if ((mask & 1) == 0) {                                      \
    392                /* ECI says skip this beat */                           \
    393                continue;                                               \
    394            }                                                           \
    395            addr = base + off[beat] * 4;                                \
    396            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    397            for (e = 0; e < 4; e++, data >>= 8) {                       \
    398                uint8_t *qd = (uint8_t *)aa32_vfp_qreg(env, qnidx + e); \
    399                qd[H1(off[beat])] = data;                               \
    400            }                                                           \
    401        }                                                               \
    402    }
    403
    404#define DO_VLD4H(OP, O1, O2)                                            \
    405    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    406                          uint32_t base)                                \
    407    {                                                                   \
    408        int beat;                                                       \
    409        uint16_t mask = mve_eci_mask(env);                              \
    410        static const uint8_t off[4] = { O1, O1, O2, O2 };               \
    411        uint32_t addr, data;                                            \
    412        int y; /* y counts 0 2 0 2 */                                   \
    413        uint16_t *qd;                                                   \
    414        for (beat = 0, y = 0; beat < 4; beat++, mask >>= 4, y ^= 2) {   \
    415            if ((mask & 1) == 0) {                                      \
    416                /* ECI says skip this beat */                           \
    417                continue;                                               \
    418            }                                                           \
    419            addr = base + off[beat] * 8 + (beat & 1) * 4;               \
    420            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    421            qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + y);             \
    422            qd[H2(off[beat])] = data;                                   \
    423            data >>= 16;                                                \
    424            qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + y + 1);         \
    425            qd[H2(off[beat])] = data;                                   \
    426        }                                                               \
    427    }
    428
    429#define DO_VLD4W(OP, O1, O2, O3, O4)                                    \
    430    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    431                          uint32_t base)                                \
    432    {                                                                   \
    433        int beat;                                                       \
    434        uint16_t mask = mve_eci_mask(env);                              \
    435        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    436        uint32_t addr, data;                                            \
    437        uint32_t *qd;                                                   \
    438        int y;                                                          \
    439        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    440            if ((mask & 1) == 0) {                                      \
    441                /* ECI says skip this beat */                           \
    442                continue;                                               \
    443            }                                                           \
    444            addr = base + off[beat] * 4;                                \
    445            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    446            y = (beat + (O1 & 2)) & 3;                                  \
    447            qd = (uint32_t *)aa32_vfp_qreg(env, qnidx + y);             \
    448            qd[H4(off[beat] >> 2)] = data;                              \
    449        }                                                               \
    450    }
    451
    452DO_VLD4B(vld40b, 0, 1, 10, 11)
    453DO_VLD4B(vld41b, 2, 3, 12, 13)
    454DO_VLD4B(vld42b, 4, 5, 14, 15)
    455DO_VLD4B(vld43b, 6, 7, 8, 9)
    456
    457DO_VLD4H(vld40h, 0, 5)
    458DO_VLD4H(vld41h, 1, 6)
    459DO_VLD4H(vld42h, 2, 7)
    460DO_VLD4H(vld43h, 3, 4)
    461
    462DO_VLD4W(vld40w, 0, 1, 10, 11)
    463DO_VLD4W(vld41w, 2, 3, 12, 13)
    464DO_VLD4W(vld42w, 4, 5, 14, 15)
    465DO_VLD4W(vld43w, 6, 7, 8, 9)
    466
    467#define DO_VLD2B(OP, O1, O2, O3, O4)                                    \
    468    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    469                          uint32_t base)                                \
    470    {                                                                   \
    471        int beat, e;                                                    \
    472        uint16_t mask = mve_eci_mask(env);                              \
    473        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    474        uint32_t addr, data;                                            \
    475        uint8_t *qd;                                                    \
    476        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    477            if ((mask & 1) == 0) {                                      \
    478                /* ECI says skip this beat */                           \
    479                continue;                                               \
    480            }                                                           \
    481            addr = base + off[beat] * 2;                                \
    482            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    483            for (e = 0; e < 4; e++, data >>= 8) {                       \
    484                qd = (uint8_t *)aa32_vfp_qreg(env, qnidx + (e & 1));    \
    485                qd[H1(off[beat] + (e >> 1))] = data;                    \
    486            }                                                           \
    487        }                                                               \
    488    }
    489
    490#define DO_VLD2H(OP, O1, O2, O3, O4)                                    \
    491    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    492                          uint32_t base)                                \
    493    {                                                                   \
    494        int beat;                                                       \
    495        uint16_t mask = mve_eci_mask(env);                              \
    496        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    497        uint32_t addr, data;                                            \
    498        int e;                                                          \
    499        uint16_t *qd;                                                   \
    500        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    501            if ((mask & 1) == 0) {                                      \
    502                /* ECI says skip this beat */                           \
    503                continue;                                               \
    504            }                                                           \
    505            addr = base + off[beat] * 4;                                \
    506            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    507            for (e = 0; e < 2; e++, data >>= 16) {                      \
    508                qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + e);         \
    509                qd[H2(off[beat])] = data;                               \
    510            }                                                           \
    511        }                                                               \
    512    }
    513
    514#define DO_VLD2W(OP, O1, O2, O3, O4)                                    \
    515    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    516                          uint32_t base)                                \
    517    {                                                                   \
    518        int beat;                                                       \
    519        uint16_t mask = mve_eci_mask(env);                              \
    520        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    521        uint32_t addr, data;                                            \
    522        uint32_t *qd;                                                   \
    523        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    524            if ((mask & 1) == 0) {                                      \
    525                /* ECI says skip this beat */                           \
    526                continue;                                               \
    527            }                                                           \
    528            addr = base + off[beat];                                    \
    529            data = cpu_ldl_le_data_ra(env, addr, GETPC());              \
    530            qd = (uint32_t *)aa32_vfp_qreg(env, qnidx + (beat & 1));    \
    531            qd[H4(off[beat] >> 3)] = data;                              \
    532        }                                                               \
    533    }
    534
    535DO_VLD2B(vld20b, 0, 2, 12, 14)
    536DO_VLD2B(vld21b, 4, 6, 8, 10)
    537
    538DO_VLD2H(vld20h, 0, 1, 6, 7)
    539DO_VLD2H(vld21h, 2, 3, 4, 5)
    540
    541DO_VLD2W(vld20w, 0, 4, 24, 28)
    542DO_VLD2W(vld21w, 8, 12, 16, 20)
    543
    544#define DO_VST4B(OP, O1, O2, O3, O4)                                    \
    545    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    546                          uint32_t base)                                \
    547    {                                                                   \
    548        int beat, e;                                                    \
    549        uint16_t mask = mve_eci_mask(env);                              \
    550        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    551        uint32_t addr, data;                                            \
    552        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    553            if ((mask & 1) == 0) {                                      \
    554                /* ECI says skip this beat */                           \
    555                continue;                                               \
    556            }                                                           \
    557            addr = base + off[beat] * 4;                                \
    558            data = 0;                                                   \
    559            for (e = 3; e >= 0; e--) {                                  \
    560                uint8_t *qd = (uint8_t *)aa32_vfp_qreg(env, qnidx + e); \
    561                data = (data << 8) | qd[H1(off[beat])];                 \
    562            }                                                           \
    563            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    564        }                                                               \
    565    }
    566
    567#define DO_VST4H(OP, O1, O2)                                            \
    568    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    569                          uint32_t base)                                \
    570    {                                                                   \
    571        int beat;                                                       \
    572        uint16_t mask = mve_eci_mask(env);                              \
    573        static const uint8_t off[4] = { O1, O1, O2, O2 };               \
    574        uint32_t addr, data;                                            \
    575        int y; /* y counts 0 2 0 2 */                                   \
    576        uint16_t *qd;                                                   \
    577        for (beat = 0, y = 0; beat < 4; beat++, mask >>= 4, y ^= 2) {   \
    578            if ((mask & 1) == 0) {                                      \
    579                /* ECI says skip this beat */                           \
    580                continue;                                               \
    581            }                                                           \
    582            addr = base + off[beat] * 8 + (beat & 1) * 4;               \
    583            qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + y);             \
    584            data = qd[H2(off[beat])];                                   \
    585            qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + y + 1);         \
    586            data |= qd[H2(off[beat])] << 16;                            \
    587            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    588        }                                                               \
    589    }
    590
    591#define DO_VST4W(OP, O1, O2, O3, O4)                                    \
    592    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    593                          uint32_t base)                                \
    594    {                                                                   \
    595        int beat;                                                       \
    596        uint16_t mask = mve_eci_mask(env);                              \
    597        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    598        uint32_t addr, data;                                            \
    599        uint32_t *qd;                                                   \
    600        int y;                                                          \
    601        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    602            if ((mask & 1) == 0) {                                      \
    603                /* ECI says skip this beat */                           \
    604                continue;                                               \
    605            }                                                           \
    606            addr = base + off[beat] * 4;                                \
    607            y = (beat + (O1 & 2)) & 3;                                  \
    608            qd = (uint32_t *)aa32_vfp_qreg(env, qnidx + y);             \
    609            data = qd[H4(off[beat] >> 2)];                              \
    610            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    611        }                                                               \
    612    }
    613
    614DO_VST4B(vst40b, 0, 1, 10, 11)
    615DO_VST4B(vst41b, 2, 3, 12, 13)
    616DO_VST4B(vst42b, 4, 5, 14, 15)
    617DO_VST4B(vst43b, 6, 7, 8, 9)
    618
    619DO_VST4H(vst40h, 0, 5)
    620DO_VST4H(vst41h, 1, 6)
    621DO_VST4H(vst42h, 2, 7)
    622DO_VST4H(vst43h, 3, 4)
    623
    624DO_VST4W(vst40w, 0, 1, 10, 11)
    625DO_VST4W(vst41w, 2, 3, 12, 13)
    626DO_VST4W(vst42w, 4, 5, 14, 15)
    627DO_VST4W(vst43w, 6, 7, 8, 9)
    628
    629#define DO_VST2B(OP, O1, O2, O3, O4)                                    \
    630    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    631                          uint32_t base)                                \
    632    {                                                                   \
    633        int beat, e;                                                    \
    634        uint16_t mask = mve_eci_mask(env);                              \
    635        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    636        uint32_t addr, data;                                            \
    637        uint8_t *qd;                                                    \
    638        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    639            if ((mask & 1) == 0) {                                      \
    640                /* ECI says skip this beat */                           \
    641                continue;                                               \
    642            }                                                           \
    643            addr = base + off[beat] * 2;                                \
    644            data = 0;                                                   \
    645            for (e = 3; e >= 0; e--) {                                  \
    646                qd = (uint8_t *)aa32_vfp_qreg(env, qnidx + (e & 1));    \
    647                data = (data << 8) | qd[H1(off[beat] + (e >> 1))];      \
    648            }                                                           \
    649            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    650        }                                                               \
    651    }
    652
    653#define DO_VST2H(OP, O1, O2, O3, O4)                                    \
    654    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    655                          uint32_t base)                                \
    656    {                                                                   \
    657        int beat;                                                       \
    658        uint16_t mask = mve_eci_mask(env);                              \
    659        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    660        uint32_t addr, data;                                            \
    661        int e;                                                          \
    662        uint16_t *qd;                                                   \
    663        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    664            if ((mask & 1) == 0) {                                      \
    665                /* ECI says skip this beat */                           \
    666                continue;                                               \
    667            }                                                           \
    668            addr = base + off[beat] * 4;                                \
    669            data = 0;                                                   \
    670            for (e = 1; e >= 0; e--) {                                  \
    671                qd = (uint16_t *)aa32_vfp_qreg(env, qnidx + e);         \
    672                data = (data << 16) | qd[H2(off[beat])];                \
    673            }                                                           \
    674            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    675        }                                                               \
    676    }
    677
    678#define DO_VST2W(OP, O1, O2, O3, O4)                                    \
    679    void HELPER(mve_##OP)(CPUARMState *env, uint32_t qnidx,             \
    680                          uint32_t base)                                \
    681    {                                                                   \
    682        int beat;                                                       \
    683        uint16_t mask = mve_eci_mask(env);                              \
    684        static const uint8_t off[4] = { O1, O2, O3, O4 };               \
    685        uint32_t addr, data;                                            \
    686        uint32_t *qd;                                                   \
    687        for (beat = 0; beat < 4; beat++, mask >>= 4) {                  \
    688            if ((mask & 1) == 0) {                                      \
    689                /* ECI says skip this beat */                           \
    690                continue;                                               \
    691            }                                                           \
    692            addr = base + off[beat];                                    \
    693            qd = (uint32_t *)aa32_vfp_qreg(env, qnidx + (beat & 1));    \
    694            data = qd[H4(off[beat] >> 3)];                              \
    695            cpu_stl_le_data_ra(env, addr, data, GETPC());               \
    696        }                                                               \
    697    }
    698
    699DO_VST2B(vst20b, 0, 2, 12, 14)
    700DO_VST2B(vst21b, 4, 6, 8, 10)
    701
    702DO_VST2H(vst20h, 0, 1, 6, 7)
    703DO_VST2H(vst21h, 2, 3, 4, 5)
    704
    705DO_VST2W(vst20w, 0, 4, 24, 28)
    706DO_VST2W(vst21w, 8, 12, 16, 20)
    707
    708/*
    709 * The mergemask(D, R, M) macro performs the operation "*D = R" but
    710 * storing only the bytes which correspond to 1 bits in M,
    711 * leaving other bytes in *D unchanged. We use _Generic
    712 * to select the correct implementation based on the type of D.
    713 */
    714
    715static void mergemask_ub(uint8_t *d, uint8_t r, uint16_t mask)
    716{
    717    if (mask & 1) {
    718        *d = r;
    719    }
    720}
    721
    722static void mergemask_sb(int8_t *d, int8_t r, uint16_t mask)
    723{
    724    mergemask_ub((uint8_t *)d, r, mask);
    725}
    726
    727static void mergemask_uh(uint16_t *d, uint16_t r, uint16_t mask)
    728{
    729    uint16_t bmask = expand_pred_b_data[mask & 3];
    730    *d = (*d & ~bmask) | (r & bmask);
    731}
    732
    733static void mergemask_sh(int16_t *d, int16_t r, uint16_t mask)
    734{
    735    mergemask_uh((uint16_t *)d, r, mask);
    736}
    737
    738static void mergemask_uw(uint32_t *d, uint32_t r, uint16_t mask)
    739{
    740    uint32_t bmask = expand_pred_b_data[mask & 0xf];
    741    *d = (*d & ~bmask) | (r & bmask);
    742}
    743
    744static void mergemask_sw(int32_t *d, int32_t r, uint16_t mask)
    745{
    746    mergemask_uw((uint32_t *)d, r, mask);
    747}
    748
    749static void mergemask_uq(uint64_t *d, uint64_t r, uint16_t mask)
    750{
    751    uint64_t bmask = expand_pred_b_data[mask & 0xff];
    752    *d = (*d & ~bmask) | (r & bmask);
    753}
    754
    755static void mergemask_sq(int64_t *d, int64_t r, uint16_t mask)
    756{
    757    mergemask_uq((uint64_t *)d, r, mask);
    758}
    759
    760#define mergemask(D, R, M)                      \
    761    _Generic(D,                                 \
    762             uint8_t *: mergemask_ub,           \
    763             int8_t *:  mergemask_sb,           \
    764             uint16_t *: mergemask_uh,          \
    765             int16_t *:  mergemask_sh,          \
    766             uint32_t *: mergemask_uw,          \
    767             int32_t *:  mergemask_sw,          \
    768             uint64_t *: mergemask_uq,          \
    769             int64_t *:  mergemask_sq)(D, R, M)
    770
    771void HELPER(mve_vdup)(CPUARMState *env, void *vd, uint32_t val)
    772{
    773    /*
    774     * The generated code already replicated an 8 or 16 bit constant
    775     * into the 32-bit value, so we only need to write the 32-bit
    776     * value to all elements of the Qreg, allowing for predication.
    777     */
    778    uint32_t *d = vd;
    779    uint16_t mask = mve_element_mask(env);
    780    unsigned e;
    781    for (e = 0; e < 16 / 4; e++, mask >>= 4) {
    782        mergemask(&d[H4(e)], val, mask);
    783    }
    784    mve_advance_vpt(env);
    785}
    786
    787#define DO_1OP(OP, ESIZE, TYPE, FN)                                     \
    788    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm)         \
    789    {                                                                   \
    790        TYPE *d = vd, *m = vm;                                          \
    791        uint16_t mask = mve_element_mask(env);                          \
    792        unsigned e;                                                     \
    793        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
    794            mergemask(&d[H##ESIZE(e)], FN(m[H##ESIZE(e)]), mask);       \
    795        }                                                               \
    796        mve_advance_vpt(env);                                           \
    797    }
    798
    799#define DO_CLS_B(N)   (clrsb32(N) - 24)
    800#define DO_CLS_H(N)   (clrsb32(N) - 16)
    801
    802DO_1OP(vclsb, 1, int8_t, DO_CLS_B)
    803DO_1OP(vclsh, 2, int16_t, DO_CLS_H)
    804DO_1OP(vclsw, 4, int32_t, clrsb32)
    805
    806#define DO_CLZ_B(N)   (clz32(N) - 24)
    807#define DO_CLZ_H(N)   (clz32(N) - 16)
    808
    809DO_1OP(vclzb, 1, uint8_t, DO_CLZ_B)
    810DO_1OP(vclzh, 2, uint16_t, DO_CLZ_H)
    811DO_1OP(vclzw, 4, uint32_t, clz32)
    812
    813DO_1OP(vrev16b, 2, uint16_t, bswap16)
    814DO_1OP(vrev32b, 4, uint32_t, bswap32)
    815DO_1OP(vrev32h, 4, uint32_t, hswap32)
    816DO_1OP(vrev64b, 8, uint64_t, bswap64)
    817DO_1OP(vrev64h, 8, uint64_t, hswap64)
    818DO_1OP(vrev64w, 8, uint64_t, wswap64)
    819
    820#define DO_NOT(N) (~(N))
    821
    822DO_1OP(vmvn, 8, uint64_t, DO_NOT)
    823
    824#define DO_ABS(N) ((N) < 0 ? -(N) : (N))
    825#define DO_FABSH(N)  ((N) & dup_const(MO_16, 0x7fff))
    826#define DO_FABSS(N)  ((N) & dup_const(MO_32, 0x7fffffff))
    827
    828DO_1OP(vabsb, 1, int8_t, DO_ABS)
    829DO_1OP(vabsh, 2, int16_t, DO_ABS)
    830DO_1OP(vabsw, 4, int32_t, DO_ABS)
    831
    832/* We can do these 64 bits at a time */
    833DO_1OP(vfabsh, 8, uint64_t, DO_FABSH)
    834DO_1OP(vfabss, 8, uint64_t, DO_FABSS)
    835
    836#define DO_NEG(N)    (-(N))
    837#define DO_FNEGH(N) ((N) ^ dup_const(MO_16, 0x8000))
    838#define DO_FNEGS(N) ((N) ^ dup_const(MO_32, 0x80000000))
    839
    840DO_1OP(vnegb, 1, int8_t, DO_NEG)
    841DO_1OP(vnegh, 2, int16_t, DO_NEG)
    842DO_1OP(vnegw, 4, int32_t, DO_NEG)
    843
    844/* We can do these 64 bits at a time */
    845DO_1OP(vfnegh, 8, uint64_t, DO_FNEGH)
    846DO_1OP(vfnegs, 8, uint64_t, DO_FNEGS)
    847
    848/*
    849 * 1 operand immediates: Vda is destination and possibly also one source.
    850 * All these insns work at 64-bit widths.
    851 */
    852#define DO_1OP_IMM(OP, FN)                                              \
    853    void HELPER(mve_##OP)(CPUARMState *env, void *vda, uint64_t imm)    \
    854    {                                                                   \
    855        uint64_t *da = vda;                                             \
    856        uint16_t mask = mve_element_mask(env);                          \
    857        unsigned e;                                                     \
    858        for (e = 0; e < 16 / 8; e++, mask >>= 8) {                      \
    859            mergemask(&da[H8(e)], FN(da[H8(e)], imm), mask);            \
    860        }                                                               \
    861        mve_advance_vpt(env);                                           \
    862    }
    863
    864#define DO_MOVI(N, I) (I)
    865#define DO_ANDI(N, I) ((N) & (I))
    866#define DO_ORRI(N, I) ((N) | (I))
    867
    868DO_1OP_IMM(vmovi, DO_MOVI)
    869DO_1OP_IMM(vandi, DO_ANDI)
    870DO_1OP_IMM(vorri, DO_ORRI)
    871
    872#define DO_2OP(OP, ESIZE, TYPE, FN)                                     \
    873    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
    874                                void *vd, void *vn, void *vm)           \
    875    {                                                                   \
    876        TYPE *d = vd, *n = vn, *m = vm;                                 \
    877        uint16_t mask = mve_element_mask(env);                          \
    878        unsigned e;                                                     \
    879        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
    880            mergemask(&d[H##ESIZE(e)],                                  \
    881                      FN(n[H##ESIZE(e)], m[H##ESIZE(e)]), mask);        \
    882        }                                                               \
    883        mve_advance_vpt(env);                                           \
    884    }
    885
    886/* provide unsigned 2-op helpers for all sizes */
    887#define DO_2OP_U(OP, FN)                        \
    888    DO_2OP(OP##b, 1, uint8_t, FN)               \
    889    DO_2OP(OP##h, 2, uint16_t, FN)              \
    890    DO_2OP(OP##w, 4, uint32_t, FN)
    891
    892/* provide signed 2-op helpers for all sizes */
    893#define DO_2OP_S(OP, FN)                        \
    894    DO_2OP(OP##b, 1, int8_t, FN)                \
    895    DO_2OP(OP##h, 2, int16_t, FN)               \
    896    DO_2OP(OP##w, 4, int32_t, FN)
    897
    898/*
    899 * "Long" operations where two half-sized inputs (taken from either the
    900 * top or the bottom of the input vector) produce a double-width result.
    901 * Here ESIZE, TYPE are for the input, and LESIZE, LTYPE for the output.
    902 */
    903#define DO_2OP_L(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN)               \
    904    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn, void *vm) \
    905    {                                                                   \
    906        LTYPE *d = vd;                                                  \
    907        TYPE *n = vn, *m = vm;                                          \
    908        uint16_t mask = mve_element_mask(env);                          \
    909        unsigned le;                                                    \
    910        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
    911            LTYPE r = FN((LTYPE)n[H##ESIZE(le * 2 + TOP)],              \
    912                         m[H##ESIZE(le * 2 + TOP)]);                    \
    913            mergemask(&d[H##LESIZE(le)], r, mask);                      \
    914        }                                                               \
    915        mve_advance_vpt(env);                                           \
    916    }
    917
    918#define DO_2OP_SAT(OP, ESIZE, TYPE, FN)                                 \
    919    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn, void *vm) \
    920    {                                                                   \
    921        TYPE *d = vd, *n = vn, *m = vm;                                 \
    922        uint16_t mask = mve_element_mask(env);                          \
    923        unsigned e;                                                     \
    924        bool qc = false;                                                \
    925        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
    926            bool sat = false;                                           \
    927            TYPE r = FN(n[H##ESIZE(e)], m[H##ESIZE(e)], &sat);          \
    928            mergemask(&d[H##ESIZE(e)], r, mask);                        \
    929            qc |= sat & mask & 1;                                       \
    930        }                                                               \
    931        if (qc) {                                                       \
    932            env->vfp.qc[0] = qc;                                        \
    933        }                                                               \
    934        mve_advance_vpt(env);                                           \
    935    }
    936
    937/* provide unsigned 2-op helpers for all sizes */
    938#define DO_2OP_SAT_U(OP, FN)                    \
    939    DO_2OP_SAT(OP##b, 1, uint8_t, FN)           \
    940    DO_2OP_SAT(OP##h, 2, uint16_t, FN)          \
    941    DO_2OP_SAT(OP##w, 4, uint32_t, FN)
    942
    943/* provide signed 2-op helpers for all sizes */
    944#define DO_2OP_SAT_S(OP, FN)                    \
    945    DO_2OP_SAT(OP##b, 1, int8_t, FN)            \
    946    DO_2OP_SAT(OP##h, 2, int16_t, FN)           \
    947    DO_2OP_SAT(OP##w, 4, int32_t, FN)
    948
    949#define DO_AND(N, M)  ((N) & (M))
    950#define DO_BIC(N, M)  ((N) & ~(M))
    951#define DO_ORR(N, M)  ((N) | (M))
    952#define DO_ORN(N, M)  ((N) | ~(M))
    953#define DO_EOR(N, M)  ((N) ^ (M))
    954
    955DO_2OP(vand, 8, uint64_t, DO_AND)
    956DO_2OP(vbic, 8, uint64_t, DO_BIC)
    957DO_2OP(vorr, 8, uint64_t, DO_ORR)
    958DO_2OP(vorn, 8, uint64_t, DO_ORN)
    959DO_2OP(veor, 8, uint64_t, DO_EOR)
    960
    961#define DO_ADD(N, M) ((N) + (M))
    962#define DO_SUB(N, M) ((N) - (M))
    963#define DO_MUL(N, M) ((N) * (M))
    964
    965DO_2OP_U(vadd, DO_ADD)
    966DO_2OP_U(vsub, DO_SUB)
    967DO_2OP_U(vmul, DO_MUL)
    968
    969DO_2OP_L(vmullbsb, 0, 1, int8_t, 2, int16_t, DO_MUL)
    970DO_2OP_L(vmullbsh, 0, 2, int16_t, 4, int32_t, DO_MUL)
    971DO_2OP_L(vmullbsw, 0, 4, int32_t, 8, int64_t, DO_MUL)
    972DO_2OP_L(vmullbub, 0, 1, uint8_t, 2, uint16_t, DO_MUL)
    973DO_2OP_L(vmullbuh, 0, 2, uint16_t, 4, uint32_t, DO_MUL)
    974DO_2OP_L(vmullbuw, 0, 4, uint32_t, 8, uint64_t, DO_MUL)
    975
    976DO_2OP_L(vmulltsb, 1, 1, int8_t, 2, int16_t, DO_MUL)
    977DO_2OP_L(vmulltsh, 1, 2, int16_t, 4, int32_t, DO_MUL)
    978DO_2OP_L(vmulltsw, 1, 4, int32_t, 8, int64_t, DO_MUL)
    979DO_2OP_L(vmulltub, 1, 1, uint8_t, 2, uint16_t, DO_MUL)
    980DO_2OP_L(vmulltuh, 1, 2, uint16_t, 4, uint32_t, DO_MUL)
    981DO_2OP_L(vmulltuw, 1, 4, uint32_t, 8, uint64_t, DO_MUL)
    982
    983/*
    984 * Polynomial multiply. We can always do this generating 64 bits
    985 * of the result at a time, so we don't need to use DO_2OP_L.
    986 */
    987#define VMULLPH_MASK 0x00ff00ff00ff00ffULL
    988#define VMULLPW_MASK 0x0000ffff0000ffffULL
    989#define DO_VMULLPBH(N, M) pmull_h((N) & VMULLPH_MASK, (M) & VMULLPH_MASK)
    990#define DO_VMULLPTH(N, M) DO_VMULLPBH((N) >> 8, (M) >> 8)
    991#define DO_VMULLPBW(N, M) pmull_w((N) & VMULLPW_MASK, (M) & VMULLPW_MASK)
    992#define DO_VMULLPTW(N, M) DO_VMULLPBW((N) >> 16, (M) >> 16)
    993
    994DO_2OP(vmullpbh, 8, uint64_t, DO_VMULLPBH)
    995DO_2OP(vmullpth, 8, uint64_t, DO_VMULLPTH)
    996DO_2OP(vmullpbw, 8, uint64_t, DO_VMULLPBW)
    997DO_2OP(vmullptw, 8, uint64_t, DO_VMULLPTW)
    998
    999/*
   1000 * Because the computation type is at least twice as large as required,
   1001 * these work for both signed and unsigned source types.
   1002 */
   1003static inline uint8_t do_mulh_b(int32_t n, int32_t m)
   1004{
   1005    return (n * m) >> 8;
   1006}
   1007
   1008static inline uint16_t do_mulh_h(int32_t n, int32_t m)
   1009{
   1010    return (n * m) >> 16;
   1011}
   1012
   1013static inline uint32_t do_mulh_w(int64_t n, int64_t m)
   1014{
   1015    return (n * m) >> 32;
   1016}
   1017
   1018static inline uint8_t do_rmulh_b(int32_t n, int32_t m)
   1019{
   1020    return (n * m + (1U << 7)) >> 8;
   1021}
   1022
   1023static inline uint16_t do_rmulh_h(int32_t n, int32_t m)
   1024{
   1025    return (n * m + (1U << 15)) >> 16;
   1026}
   1027
   1028static inline uint32_t do_rmulh_w(int64_t n, int64_t m)
   1029{
   1030    return (n * m + (1U << 31)) >> 32;
   1031}
   1032
   1033DO_2OP(vmulhsb, 1, int8_t, do_mulh_b)
   1034DO_2OP(vmulhsh, 2, int16_t, do_mulh_h)
   1035DO_2OP(vmulhsw, 4, int32_t, do_mulh_w)
   1036DO_2OP(vmulhub, 1, uint8_t, do_mulh_b)
   1037DO_2OP(vmulhuh, 2, uint16_t, do_mulh_h)
   1038DO_2OP(vmulhuw, 4, uint32_t, do_mulh_w)
   1039
   1040DO_2OP(vrmulhsb, 1, int8_t, do_rmulh_b)
   1041DO_2OP(vrmulhsh, 2, int16_t, do_rmulh_h)
   1042DO_2OP(vrmulhsw, 4, int32_t, do_rmulh_w)
   1043DO_2OP(vrmulhub, 1, uint8_t, do_rmulh_b)
   1044DO_2OP(vrmulhuh, 2, uint16_t, do_rmulh_h)
   1045DO_2OP(vrmulhuw, 4, uint32_t, do_rmulh_w)
   1046
   1047#define DO_MAX(N, M)  ((N) >= (M) ? (N) : (M))
   1048#define DO_MIN(N, M)  ((N) >= (M) ? (M) : (N))
   1049
   1050DO_2OP_S(vmaxs, DO_MAX)
   1051DO_2OP_U(vmaxu, DO_MAX)
   1052DO_2OP_S(vmins, DO_MIN)
   1053DO_2OP_U(vminu, DO_MIN)
   1054
   1055#define DO_ABD(N, M)  ((N) >= (M) ? (N) - (M) : (M) - (N))
   1056
   1057DO_2OP_S(vabds, DO_ABD)
   1058DO_2OP_U(vabdu, DO_ABD)
   1059
   1060static inline uint32_t do_vhadd_u(uint32_t n, uint32_t m)
   1061{
   1062    return ((uint64_t)n + m) >> 1;
   1063}
   1064
   1065static inline int32_t do_vhadd_s(int32_t n, int32_t m)
   1066{
   1067    return ((int64_t)n + m) >> 1;
   1068}
   1069
   1070static inline uint32_t do_vhsub_u(uint32_t n, uint32_t m)
   1071{
   1072    return ((uint64_t)n - m) >> 1;
   1073}
   1074
   1075static inline int32_t do_vhsub_s(int32_t n, int32_t m)
   1076{
   1077    return ((int64_t)n - m) >> 1;
   1078}
   1079
   1080DO_2OP_S(vhadds, do_vhadd_s)
   1081DO_2OP_U(vhaddu, do_vhadd_u)
   1082DO_2OP_S(vhsubs, do_vhsub_s)
   1083DO_2OP_U(vhsubu, do_vhsub_u)
   1084
   1085#define DO_VSHLS(N, M) do_sqrshl_bhs(N, (int8_t)(M), sizeof(N) * 8, false, NULL)
   1086#define DO_VSHLU(N, M) do_uqrshl_bhs(N, (int8_t)(M), sizeof(N) * 8, false, NULL)
   1087#define DO_VRSHLS(N, M) do_sqrshl_bhs(N, (int8_t)(M), sizeof(N) * 8, true, NULL)
   1088#define DO_VRSHLU(N, M) do_uqrshl_bhs(N, (int8_t)(M), sizeof(N) * 8, true, NULL)
   1089
   1090DO_2OP_S(vshls, DO_VSHLS)
   1091DO_2OP_U(vshlu, DO_VSHLU)
   1092DO_2OP_S(vrshls, DO_VRSHLS)
   1093DO_2OP_U(vrshlu, DO_VRSHLU)
   1094
   1095#define DO_RHADD_S(N, M) (((int64_t)(N) + (M) + 1) >> 1)
   1096#define DO_RHADD_U(N, M) (((uint64_t)(N) + (M) + 1) >> 1)
   1097
   1098DO_2OP_S(vrhadds, DO_RHADD_S)
   1099DO_2OP_U(vrhaddu, DO_RHADD_U)
   1100
   1101static void do_vadc(CPUARMState *env, uint32_t *d, uint32_t *n, uint32_t *m,
   1102                    uint32_t inv, uint32_t carry_in, bool update_flags)
   1103{
   1104    uint16_t mask = mve_element_mask(env);
   1105    unsigned e;
   1106
   1107    /* If any additions trigger, we will update flags. */
   1108    if (mask & 0x1111) {
   1109        update_flags = true;
   1110    }
   1111
   1112    for (e = 0; e < 16 / 4; e++, mask >>= 4) {
   1113        uint64_t r = carry_in;
   1114        r += n[H4(e)];
   1115        r += m[H4(e)] ^ inv;
   1116        if (mask & 1) {
   1117            carry_in = r >> 32;
   1118        }
   1119        mergemask(&d[H4(e)], r, mask);
   1120    }
   1121
   1122    if (update_flags) {
   1123        /* Store C, clear NZV. */
   1124        env->vfp.xregs[ARM_VFP_FPSCR] &= ~FPCR_NZCV_MASK;
   1125        env->vfp.xregs[ARM_VFP_FPSCR] |= carry_in * FPCR_C;
   1126    }
   1127    mve_advance_vpt(env);
   1128}
   1129
   1130void HELPER(mve_vadc)(CPUARMState *env, void *vd, void *vn, void *vm)
   1131{
   1132    bool carry_in = env->vfp.xregs[ARM_VFP_FPSCR] & FPCR_C;
   1133    do_vadc(env, vd, vn, vm, 0, carry_in, false);
   1134}
   1135
   1136void HELPER(mve_vsbc)(CPUARMState *env, void *vd, void *vn, void *vm)
   1137{
   1138    bool carry_in = env->vfp.xregs[ARM_VFP_FPSCR] & FPCR_C;
   1139    do_vadc(env, vd, vn, vm, -1, carry_in, false);
   1140}
   1141
   1142
   1143void HELPER(mve_vadci)(CPUARMState *env, void *vd, void *vn, void *vm)
   1144{
   1145    do_vadc(env, vd, vn, vm, 0, 0, true);
   1146}
   1147
   1148void HELPER(mve_vsbci)(CPUARMState *env, void *vd, void *vn, void *vm)
   1149{
   1150    do_vadc(env, vd, vn, vm, -1, 1, true);
   1151}
   1152
   1153#define DO_VCADD(OP, ESIZE, TYPE, FN0, FN1)                             \
   1154    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn, void *vm) \
   1155    {                                                                   \
   1156        TYPE *d = vd, *n = vn, *m = vm;                                 \
   1157        uint16_t mask = mve_element_mask(env);                          \
   1158        unsigned e;                                                     \
   1159        TYPE r[16 / ESIZE];                                             \
   1160        /* Calculate all results first to avoid overwriting inputs */   \
   1161        for (e = 0; e < 16 / ESIZE; e++) {                              \
   1162            if (!(e & 1)) {                                             \
   1163                r[e] = FN0(n[H##ESIZE(e)], m[H##ESIZE(e + 1)]);         \
   1164            } else {                                                    \
   1165                r[e] = FN1(n[H##ESIZE(e)], m[H##ESIZE(e - 1)]);         \
   1166            }                                                           \
   1167        }                                                               \
   1168        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1169            mergemask(&d[H##ESIZE(e)], r[e], mask);                     \
   1170        }                                                               \
   1171        mve_advance_vpt(env);                                           \
   1172    }
   1173
   1174#define DO_VCADD_ALL(OP, FN0, FN1)              \
   1175    DO_VCADD(OP##b, 1, int8_t, FN0, FN1)        \
   1176    DO_VCADD(OP##h, 2, int16_t, FN0, FN1)       \
   1177    DO_VCADD(OP##w, 4, int32_t, FN0, FN1)
   1178
   1179DO_VCADD_ALL(vcadd90, DO_SUB, DO_ADD)
   1180DO_VCADD_ALL(vcadd270, DO_ADD, DO_SUB)
   1181DO_VCADD_ALL(vhcadd90, do_vhsub_s, do_vhadd_s)
   1182DO_VCADD_ALL(vhcadd270, do_vhadd_s, do_vhsub_s)
   1183
   1184static inline int32_t do_sat_bhw(int64_t val, int64_t min, int64_t max, bool *s)
   1185{
   1186    if (val > max) {
   1187        *s = true;
   1188        return max;
   1189    } else if (val < min) {
   1190        *s = true;
   1191        return min;
   1192    }
   1193    return val;
   1194}
   1195
   1196#define DO_SQADD_B(n, m, s) do_sat_bhw((int64_t)n + m, INT8_MIN, INT8_MAX, s)
   1197#define DO_SQADD_H(n, m, s) do_sat_bhw((int64_t)n + m, INT16_MIN, INT16_MAX, s)
   1198#define DO_SQADD_W(n, m, s) do_sat_bhw((int64_t)n + m, INT32_MIN, INT32_MAX, s)
   1199
   1200#define DO_UQADD_B(n, m, s) do_sat_bhw((int64_t)n + m, 0, UINT8_MAX, s)
   1201#define DO_UQADD_H(n, m, s) do_sat_bhw((int64_t)n + m, 0, UINT16_MAX, s)
   1202#define DO_UQADD_W(n, m, s) do_sat_bhw((int64_t)n + m, 0, UINT32_MAX, s)
   1203
   1204#define DO_SQSUB_B(n, m, s) do_sat_bhw((int64_t)n - m, INT8_MIN, INT8_MAX, s)
   1205#define DO_SQSUB_H(n, m, s) do_sat_bhw((int64_t)n - m, INT16_MIN, INT16_MAX, s)
   1206#define DO_SQSUB_W(n, m, s) do_sat_bhw((int64_t)n - m, INT32_MIN, INT32_MAX, s)
   1207
   1208#define DO_UQSUB_B(n, m, s) do_sat_bhw((int64_t)n - m, 0, UINT8_MAX, s)
   1209#define DO_UQSUB_H(n, m, s) do_sat_bhw((int64_t)n - m, 0, UINT16_MAX, s)
   1210#define DO_UQSUB_W(n, m, s) do_sat_bhw((int64_t)n - m, 0, UINT32_MAX, s)
   1211
   1212/*
   1213 * For QDMULH and QRDMULH we simplify "double and shift by esize" into
   1214 * "shift by esize-1", adjusting the QRDMULH rounding constant to match.
   1215 */
   1216#define DO_QDMULH_B(n, m, s) do_sat_bhw(((int64_t)n * m) >> 7, \
   1217                                        INT8_MIN, INT8_MAX, s)
   1218#define DO_QDMULH_H(n, m, s) do_sat_bhw(((int64_t)n * m) >> 15, \
   1219                                        INT16_MIN, INT16_MAX, s)
   1220#define DO_QDMULH_W(n, m, s) do_sat_bhw(((int64_t)n * m) >> 31, \
   1221                                        INT32_MIN, INT32_MAX, s)
   1222
   1223#define DO_QRDMULH_B(n, m, s) do_sat_bhw(((int64_t)n * m + (1 << 6)) >> 7, \
   1224                                         INT8_MIN, INT8_MAX, s)
   1225#define DO_QRDMULH_H(n, m, s) do_sat_bhw(((int64_t)n * m + (1 << 14)) >> 15, \
   1226                                         INT16_MIN, INT16_MAX, s)
   1227#define DO_QRDMULH_W(n, m, s) do_sat_bhw(((int64_t)n * m + (1 << 30)) >> 31, \
   1228                                         INT32_MIN, INT32_MAX, s)
   1229
   1230DO_2OP_SAT(vqdmulhb, 1, int8_t, DO_QDMULH_B)
   1231DO_2OP_SAT(vqdmulhh, 2, int16_t, DO_QDMULH_H)
   1232DO_2OP_SAT(vqdmulhw, 4, int32_t, DO_QDMULH_W)
   1233
   1234DO_2OP_SAT(vqrdmulhb, 1, int8_t, DO_QRDMULH_B)
   1235DO_2OP_SAT(vqrdmulhh, 2, int16_t, DO_QRDMULH_H)
   1236DO_2OP_SAT(vqrdmulhw, 4, int32_t, DO_QRDMULH_W)
   1237
   1238DO_2OP_SAT(vqaddub, 1, uint8_t, DO_UQADD_B)
   1239DO_2OP_SAT(vqadduh, 2, uint16_t, DO_UQADD_H)
   1240DO_2OP_SAT(vqadduw, 4, uint32_t, DO_UQADD_W)
   1241DO_2OP_SAT(vqaddsb, 1, int8_t, DO_SQADD_B)
   1242DO_2OP_SAT(vqaddsh, 2, int16_t, DO_SQADD_H)
   1243DO_2OP_SAT(vqaddsw, 4, int32_t, DO_SQADD_W)
   1244
   1245DO_2OP_SAT(vqsubub, 1, uint8_t, DO_UQSUB_B)
   1246DO_2OP_SAT(vqsubuh, 2, uint16_t, DO_UQSUB_H)
   1247DO_2OP_SAT(vqsubuw, 4, uint32_t, DO_UQSUB_W)
   1248DO_2OP_SAT(vqsubsb, 1, int8_t, DO_SQSUB_B)
   1249DO_2OP_SAT(vqsubsh, 2, int16_t, DO_SQSUB_H)
   1250DO_2OP_SAT(vqsubsw, 4, int32_t, DO_SQSUB_W)
   1251
   1252/*
   1253 * This wrapper fixes up the impedance mismatch between do_sqrshl_bhs()
   1254 * and friends wanting a uint32_t* sat and our needing a bool*.
   1255 */
   1256#define WRAP_QRSHL_HELPER(FN, N, M, ROUND, satp)                        \
   1257    ({                                                                  \
   1258        uint32_t su32 = 0;                                              \
   1259        typeof(N) r = FN(N, (int8_t)(M), sizeof(N) * 8, ROUND, &su32);  \
   1260        if (su32) {                                                     \
   1261            *satp = true;                                               \
   1262        }                                                               \
   1263        r;                                                              \
   1264    })
   1265
   1266#define DO_SQSHL_OP(N, M, satp) \
   1267    WRAP_QRSHL_HELPER(do_sqrshl_bhs, N, M, false, satp)
   1268#define DO_UQSHL_OP(N, M, satp) \
   1269    WRAP_QRSHL_HELPER(do_uqrshl_bhs, N, M, false, satp)
   1270#define DO_SQRSHL_OP(N, M, satp) \
   1271    WRAP_QRSHL_HELPER(do_sqrshl_bhs, N, M, true, satp)
   1272#define DO_UQRSHL_OP(N, M, satp) \
   1273    WRAP_QRSHL_HELPER(do_uqrshl_bhs, N, M, true, satp)
   1274#define DO_SUQSHL_OP(N, M, satp) \
   1275    WRAP_QRSHL_HELPER(do_suqrshl_bhs, N, M, false, satp)
   1276
   1277DO_2OP_SAT_S(vqshls, DO_SQSHL_OP)
   1278DO_2OP_SAT_U(vqshlu, DO_UQSHL_OP)
   1279DO_2OP_SAT_S(vqrshls, DO_SQRSHL_OP)
   1280DO_2OP_SAT_U(vqrshlu, DO_UQRSHL_OP)
   1281
   1282/*
   1283 * Multiply add dual returning high half
   1284 * The 'FN' here takes four inputs A, B, C, D, a 0/1 indicator of
   1285 * whether to add the rounding constant, and the pointer to the
   1286 * saturation flag, and should do "(A * B + C * D) * 2 + rounding constant",
   1287 * saturate to twice the input size and return the high half; or
   1288 * (A * B - C * D) etc for VQDMLSDH.
   1289 */
   1290#define DO_VQDMLADH_OP(OP, ESIZE, TYPE, XCHG, ROUND, FN)                \
   1291    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1292                                void *vm)                               \
   1293    {                                                                   \
   1294        TYPE *d = vd, *n = vn, *m = vm;                                 \
   1295        uint16_t mask = mve_element_mask(env);                          \
   1296        unsigned e;                                                     \
   1297        bool qc = false;                                                \
   1298        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1299            bool sat = false;                                           \
   1300            if ((e & 1) == XCHG) {                                      \
   1301                TYPE r = FN(n[H##ESIZE(e)],                             \
   1302                            m[H##ESIZE(e - XCHG)],                      \
   1303                            n[H##ESIZE(e + (1 - 2 * XCHG))],            \
   1304                            m[H##ESIZE(e + (1 - XCHG))],                \
   1305                            ROUND, &sat);                               \
   1306                mergemask(&d[H##ESIZE(e)], r, mask);                    \
   1307                qc |= sat & mask & 1;                                   \
   1308            }                                                           \
   1309        }                                                               \
   1310        if (qc) {                                                       \
   1311            env->vfp.qc[0] = qc;                                        \
   1312        }                                                               \
   1313        mve_advance_vpt(env);                                           \
   1314    }
   1315
   1316static int8_t do_vqdmladh_b(int8_t a, int8_t b, int8_t c, int8_t d,
   1317                            int round, bool *sat)
   1318{
   1319    int64_t r = ((int64_t)a * b + (int64_t)c * d) * 2 + (round << 7);
   1320    return do_sat_bhw(r, INT16_MIN, INT16_MAX, sat) >> 8;
   1321}
   1322
   1323static int16_t do_vqdmladh_h(int16_t a, int16_t b, int16_t c, int16_t d,
   1324                             int round, bool *sat)
   1325{
   1326    int64_t r = ((int64_t)a * b + (int64_t)c * d) * 2 + (round << 15);
   1327    return do_sat_bhw(r, INT32_MIN, INT32_MAX, sat) >> 16;
   1328}
   1329
   1330static int32_t do_vqdmladh_w(int32_t a, int32_t b, int32_t c, int32_t d,
   1331                             int round, bool *sat)
   1332{
   1333    int64_t m1 = (int64_t)a * b;
   1334    int64_t m2 = (int64_t)c * d;
   1335    int64_t r;
   1336    /*
   1337     * Architecturally we should do the entire add, double, round
   1338     * and then check for saturation. We do three saturating adds,
   1339     * but we need to be careful about the order. If the first
   1340     * m1 + m2 saturates then it's impossible for the *2+rc to
   1341     * bring it back into the non-saturated range. However, if
   1342     * m1 + m2 is negative then it's possible that doing the doubling
   1343     * would take the intermediate result below INT64_MAX and the
   1344     * addition of the rounding constant then brings it back in range.
   1345     * So we add half the rounding constant before doubling rather
   1346     * than adding the rounding constant after the doubling.
   1347     */
   1348    if (sadd64_overflow(m1, m2, &r) ||
   1349        sadd64_overflow(r, (round << 30), &r) ||
   1350        sadd64_overflow(r, r, &r)) {
   1351        *sat = true;
   1352        return r < 0 ? INT32_MAX : INT32_MIN;
   1353    }
   1354    return r >> 32;
   1355}
   1356
   1357static int8_t do_vqdmlsdh_b(int8_t a, int8_t b, int8_t c, int8_t d,
   1358                            int round, bool *sat)
   1359{
   1360    int64_t r = ((int64_t)a * b - (int64_t)c * d) * 2 + (round << 7);
   1361    return do_sat_bhw(r, INT16_MIN, INT16_MAX, sat) >> 8;
   1362}
   1363
   1364static int16_t do_vqdmlsdh_h(int16_t a, int16_t b, int16_t c, int16_t d,
   1365                             int round, bool *sat)
   1366{
   1367    int64_t r = ((int64_t)a * b - (int64_t)c * d) * 2 + (round << 15);
   1368    return do_sat_bhw(r, INT32_MIN, INT32_MAX, sat) >> 16;
   1369}
   1370
   1371static int32_t do_vqdmlsdh_w(int32_t a, int32_t b, int32_t c, int32_t d,
   1372                             int round, bool *sat)
   1373{
   1374    int64_t m1 = (int64_t)a * b;
   1375    int64_t m2 = (int64_t)c * d;
   1376    int64_t r;
   1377    /* The same ordering issue as in do_vqdmladh_w applies here too */
   1378    if (ssub64_overflow(m1, m2, &r) ||
   1379        sadd64_overflow(r, (round << 30), &r) ||
   1380        sadd64_overflow(r, r, &r)) {
   1381        *sat = true;
   1382        return r < 0 ? INT32_MAX : INT32_MIN;
   1383    }
   1384    return r >> 32;
   1385}
   1386
   1387DO_VQDMLADH_OP(vqdmladhb, 1, int8_t, 0, 0, do_vqdmladh_b)
   1388DO_VQDMLADH_OP(vqdmladhh, 2, int16_t, 0, 0, do_vqdmladh_h)
   1389DO_VQDMLADH_OP(vqdmladhw, 4, int32_t, 0, 0, do_vqdmladh_w)
   1390DO_VQDMLADH_OP(vqdmladhxb, 1, int8_t, 1, 0, do_vqdmladh_b)
   1391DO_VQDMLADH_OP(vqdmladhxh, 2, int16_t, 1, 0, do_vqdmladh_h)
   1392DO_VQDMLADH_OP(vqdmladhxw, 4, int32_t, 1, 0, do_vqdmladh_w)
   1393
   1394DO_VQDMLADH_OP(vqrdmladhb, 1, int8_t, 0, 1, do_vqdmladh_b)
   1395DO_VQDMLADH_OP(vqrdmladhh, 2, int16_t, 0, 1, do_vqdmladh_h)
   1396DO_VQDMLADH_OP(vqrdmladhw, 4, int32_t, 0, 1, do_vqdmladh_w)
   1397DO_VQDMLADH_OP(vqrdmladhxb, 1, int8_t, 1, 1, do_vqdmladh_b)
   1398DO_VQDMLADH_OP(vqrdmladhxh, 2, int16_t, 1, 1, do_vqdmladh_h)
   1399DO_VQDMLADH_OP(vqrdmladhxw, 4, int32_t, 1, 1, do_vqdmladh_w)
   1400
   1401DO_VQDMLADH_OP(vqdmlsdhb, 1, int8_t, 0, 0, do_vqdmlsdh_b)
   1402DO_VQDMLADH_OP(vqdmlsdhh, 2, int16_t, 0, 0, do_vqdmlsdh_h)
   1403DO_VQDMLADH_OP(vqdmlsdhw, 4, int32_t, 0, 0, do_vqdmlsdh_w)
   1404DO_VQDMLADH_OP(vqdmlsdhxb, 1, int8_t, 1, 0, do_vqdmlsdh_b)
   1405DO_VQDMLADH_OP(vqdmlsdhxh, 2, int16_t, 1, 0, do_vqdmlsdh_h)
   1406DO_VQDMLADH_OP(vqdmlsdhxw, 4, int32_t, 1, 0, do_vqdmlsdh_w)
   1407
   1408DO_VQDMLADH_OP(vqrdmlsdhb, 1, int8_t, 0, 1, do_vqdmlsdh_b)
   1409DO_VQDMLADH_OP(vqrdmlsdhh, 2, int16_t, 0, 1, do_vqdmlsdh_h)
   1410DO_VQDMLADH_OP(vqrdmlsdhw, 4, int32_t, 0, 1, do_vqdmlsdh_w)
   1411DO_VQDMLADH_OP(vqrdmlsdhxb, 1, int8_t, 1, 1, do_vqdmlsdh_b)
   1412DO_VQDMLADH_OP(vqrdmlsdhxh, 2, int16_t, 1, 1, do_vqdmlsdh_h)
   1413DO_VQDMLADH_OP(vqrdmlsdhxw, 4, int32_t, 1, 1, do_vqdmlsdh_w)
   1414
   1415#define DO_2OP_SCALAR(OP, ESIZE, TYPE, FN)                              \
   1416    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1417                                uint32_t rm)                            \
   1418    {                                                                   \
   1419        TYPE *d = vd, *n = vn;                                          \
   1420        TYPE m = rm;                                                    \
   1421        uint16_t mask = mve_element_mask(env);                          \
   1422        unsigned e;                                                     \
   1423        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1424            mergemask(&d[H##ESIZE(e)], FN(n[H##ESIZE(e)], m), mask);    \
   1425        }                                                               \
   1426        mve_advance_vpt(env);                                           \
   1427    }
   1428
   1429#define DO_2OP_SAT_SCALAR(OP, ESIZE, TYPE, FN)                          \
   1430    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1431                                uint32_t rm)                            \
   1432    {                                                                   \
   1433        TYPE *d = vd, *n = vn;                                          \
   1434        TYPE m = rm;                                                    \
   1435        uint16_t mask = mve_element_mask(env);                          \
   1436        unsigned e;                                                     \
   1437        bool qc = false;                                                \
   1438        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1439            bool sat = false;                                           \
   1440            mergemask(&d[H##ESIZE(e)], FN(n[H##ESIZE(e)], m, &sat),     \
   1441                      mask);                                            \
   1442            qc |= sat & mask & 1;                                       \
   1443        }                                                               \
   1444        if (qc) {                                                       \
   1445            env->vfp.qc[0] = qc;                                        \
   1446        }                                                               \
   1447        mve_advance_vpt(env);                                           \
   1448    }
   1449
   1450/* "accumulating" version where FN takes d as well as n and m */
   1451#define DO_2OP_ACC_SCALAR(OP, ESIZE, TYPE, FN)                          \
   1452    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1453                                uint32_t rm)                            \
   1454    {                                                                   \
   1455        TYPE *d = vd, *n = vn;                                          \
   1456        TYPE m = rm;                                                    \
   1457        uint16_t mask = mve_element_mask(env);                          \
   1458        unsigned e;                                                     \
   1459        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1460            mergemask(&d[H##ESIZE(e)],                                  \
   1461                      FN(d[H##ESIZE(e)], n[H##ESIZE(e)], m), mask);     \
   1462        }                                                               \
   1463        mve_advance_vpt(env);                                           \
   1464    }
   1465
   1466#define DO_2OP_SAT_ACC_SCALAR(OP, ESIZE, TYPE, FN)                      \
   1467    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1468                                uint32_t rm)                            \
   1469    {                                                                   \
   1470        TYPE *d = vd, *n = vn;                                          \
   1471        TYPE m = rm;                                                    \
   1472        uint16_t mask = mve_element_mask(env);                          \
   1473        unsigned e;                                                     \
   1474        bool qc = false;                                                \
   1475        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1476            bool sat = false;                                           \
   1477            mergemask(&d[H##ESIZE(e)],                                  \
   1478                      FN(d[H##ESIZE(e)], n[H##ESIZE(e)], m, &sat),      \
   1479                      mask);                                            \
   1480            qc |= sat & mask & 1;                                       \
   1481        }                                                               \
   1482        if (qc) {                                                       \
   1483            env->vfp.qc[0] = qc;                                        \
   1484        }                                                               \
   1485        mve_advance_vpt(env);                                           \
   1486    }
   1487
   1488/* provide unsigned 2-op scalar helpers for all sizes */
   1489#define DO_2OP_SCALAR_U(OP, FN)                 \
   1490    DO_2OP_SCALAR(OP##b, 1, uint8_t, FN)        \
   1491    DO_2OP_SCALAR(OP##h, 2, uint16_t, FN)       \
   1492    DO_2OP_SCALAR(OP##w, 4, uint32_t, FN)
   1493#define DO_2OP_SCALAR_S(OP, FN)                 \
   1494    DO_2OP_SCALAR(OP##b, 1, int8_t, FN)         \
   1495    DO_2OP_SCALAR(OP##h, 2, int16_t, FN)        \
   1496    DO_2OP_SCALAR(OP##w, 4, int32_t, FN)
   1497
   1498#define DO_2OP_ACC_SCALAR_U(OP, FN)             \
   1499    DO_2OP_ACC_SCALAR(OP##b, 1, uint8_t, FN)    \
   1500    DO_2OP_ACC_SCALAR(OP##h, 2, uint16_t, FN)   \
   1501    DO_2OP_ACC_SCALAR(OP##w, 4, uint32_t, FN)
   1502
   1503DO_2OP_SCALAR_U(vadd_scalar, DO_ADD)
   1504DO_2OP_SCALAR_U(vsub_scalar, DO_SUB)
   1505DO_2OP_SCALAR_U(vmul_scalar, DO_MUL)
   1506DO_2OP_SCALAR_S(vhadds_scalar, do_vhadd_s)
   1507DO_2OP_SCALAR_U(vhaddu_scalar, do_vhadd_u)
   1508DO_2OP_SCALAR_S(vhsubs_scalar, do_vhsub_s)
   1509DO_2OP_SCALAR_U(vhsubu_scalar, do_vhsub_u)
   1510
   1511DO_2OP_SAT_SCALAR(vqaddu_scalarb, 1, uint8_t, DO_UQADD_B)
   1512DO_2OP_SAT_SCALAR(vqaddu_scalarh, 2, uint16_t, DO_UQADD_H)
   1513DO_2OP_SAT_SCALAR(vqaddu_scalarw, 4, uint32_t, DO_UQADD_W)
   1514DO_2OP_SAT_SCALAR(vqadds_scalarb, 1, int8_t, DO_SQADD_B)
   1515DO_2OP_SAT_SCALAR(vqadds_scalarh, 2, int16_t, DO_SQADD_H)
   1516DO_2OP_SAT_SCALAR(vqadds_scalarw, 4, int32_t, DO_SQADD_W)
   1517
   1518DO_2OP_SAT_SCALAR(vqsubu_scalarb, 1, uint8_t, DO_UQSUB_B)
   1519DO_2OP_SAT_SCALAR(vqsubu_scalarh, 2, uint16_t, DO_UQSUB_H)
   1520DO_2OP_SAT_SCALAR(vqsubu_scalarw, 4, uint32_t, DO_UQSUB_W)
   1521DO_2OP_SAT_SCALAR(vqsubs_scalarb, 1, int8_t, DO_SQSUB_B)
   1522DO_2OP_SAT_SCALAR(vqsubs_scalarh, 2, int16_t, DO_SQSUB_H)
   1523DO_2OP_SAT_SCALAR(vqsubs_scalarw, 4, int32_t, DO_SQSUB_W)
   1524
   1525DO_2OP_SAT_SCALAR(vqdmulh_scalarb, 1, int8_t, DO_QDMULH_B)
   1526DO_2OP_SAT_SCALAR(vqdmulh_scalarh, 2, int16_t, DO_QDMULH_H)
   1527DO_2OP_SAT_SCALAR(vqdmulh_scalarw, 4, int32_t, DO_QDMULH_W)
   1528DO_2OP_SAT_SCALAR(vqrdmulh_scalarb, 1, int8_t, DO_QRDMULH_B)
   1529DO_2OP_SAT_SCALAR(vqrdmulh_scalarh, 2, int16_t, DO_QRDMULH_H)
   1530DO_2OP_SAT_SCALAR(vqrdmulh_scalarw, 4, int32_t, DO_QRDMULH_W)
   1531
   1532static int8_t do_vqdmlah_b(int8_t a, int8_t b, int8_t c, int round, bool *sat)
   1533{
   1534    int64_t r = (int64_t)a * b * 2 + ((int64_t)c << 8) + (round << 7);
   1535    return do_sat_bhw(r, INT16_MIN, INT16_MAX, sat) >> 8;
   1536}
   1537
   1538static int16_t do_vqdmlah_h(int16_t a, int16_t b, int16_t c,
   1539                           int round, bool *sat)
   1540{
   1541    int64_t r = (int64_t)a * b * 2 + ((int64_t)c << 16) + (round << 15);
   1542    return do_sat_bhw(r, INT32_MIN, INT32_MAX, sat) >> 16;
   1543}
   1544
   1545static int32_t do_vqdmlah_w(int32_t a, int32_t b, int32_t c,
   1546                            int round, bool *sat)
   1547{
   1548    /*
   1549     * Architecturally we should do the entire add, double, round
   1550     * and then check for saturation. We do three saturating adds,
   1551     * but we need to be careful about the order. If the first
   1552     * m1 + m2 saturates then it's impossible for the *2+rc to
   1553     * bring it back into the non-saturated range. However, if
   1554     * m1 + m2 is negative then it's possible that doing the doubling
   1555     * would take the intermediate result below INT64_MAX and the
   1556     * addition of the rounding constant then brings it back in range.
   1557     * So we add half the rounding constant and half the "c << esize"
   1558     * before doubling rather than adding the rounding constant after
   1559     * the doubling.
   1560     */
   1561    int64_t m1 = (int64_t)a * b;
   1562    int64_t m2 = (int64_t)c << 31;
   1563    int64_t r;
   1564    if (sadd64_overflow(m1, m2, &r) ||
   1565        sadd64_overflow(r, (round << 30), &r) ||
   1566        sadd64_overflow(r, r, &r)) {
   1567        *sat = true;
   1568        return r < 0 ? INT32_MAX : INT32_MIN;
   1569    }
   1570    return r >> 32;
   1571}
   1572
   1573/*
   1574 * The *MLAH insns are vector * scalar + vector;
   1575 * the *MLASH insns are vector * vector + scalar
   1576 */
   1577#define DO_VQDMLAH_B(D, N, M, S) do_vqdmlah_b(N, M, D, 0, S)
   1578#define DO_VQDMLAH_H(D, N, M, S) do_vqdmlah_h(N, M, D, 0, S)
   1579#define DO_VQDMLAH_W(D, N, M, S) do_vqdmlah_w(N, M, D, 0, S)
   1580#define DO_VQRDMLAH_B(D, N, M, S) do_vqdmlah_b(N, M, D, 1, S)
   1581#define DO_VQRDMLAH_H(D, N, M, S) do_vqdmlah_h(N, M, D, 1, S)
   1582#define DO_VQRDMLAH_W(D, N, M, S) do_vqdmlah_w(N, M, D, 1, S)
   1583
   1584#define DO_VQDMLASH_B(D, N, M, S) do_vqdmlah_b(N, D, M, 0, S)
   1585#define DO_VQDMLASH_H(D, N, M, S) do_vqdmlah_h(N, D, M, 0, S)
   1586#define DO_VQDMLASH_W(D, N, M, S) do_vqdmlah_w(N, D, M, 0, S)
   1587#define DO_VQRDMLASH_B(D, N, M, S) do_vqdmlah_b(N, D, M, 1, S)
   1588#define DO_VQRDMLASH_H(D, N, M, S) do_vqdmlah_h(N, D, M, 1, S)
   1589#define DO_VQRDMLASH_W(D, N, M, S) do_vqdmlah_w(N, D, M, 1, S)
   1590
   1591DO_2OP_SAT_ACC_SCALAR(vqdmlahb, 1, int8_t, DO_VQDMLAH_B)
   1592DO_2OP_SAT_ACC_SCALAR(vqdmlahh, 2, int16_t, DO_VQDMLAH_H)
   1593DO_2OP_SAT_ACC_SCALAR(vqdmlahw, 4, int32_t, DO_VQDMLAH_W)
   1594DO_2OP_SAT_ACC_SCALAR(vqrdmlahb, 1, int8_t, DO_VQRDMLAH_B)
   1595DO_2OP_SAT_ACC_SCALAR(vqrdmlahh, 2, int16_t, DO_VQRDMLAH_H)
   1596DO_2OP_SAT_ACC_SCALAR(vqrdmlahw, 4, int32_t, DO_VQRDMLAH_W)
   1597
   1598DO_2OP_SAT_ACC_SCALAR(vqdmlashb, 1, int8_t, DO_VQDMLASH_B)
   1599DO_2OP_SAT_ACC_SCALAR(vqdmlashh, 2, int16_t, DO_VQDMLASH_H)
   1600DO_2OP_SAT_ACC_SCALAR(vqdmlashw, 4, int32_t, DO_VQDMLASH_W)
   1601DO_2OP_SAT_ACC_SCALAR(vqrdmlashb, 1, int8_t, DO_VQRDMLASH_B)
   1602DO_2OP_SAT_ACC_SCALAR(vqrdmlashh, 2, int16_t, DO_VQRDMLASH_H)
   1603DO_2OP_SAT_ACC_SCALAR(vqrdmlashw, 4, int32_t, DO_VQRDMLASH_W)
   1604
   1605/* Vector by scalar plus vector */
   1606#define DO_VMLA(D, N, M) ((N) * (M) + (D))
   1607
   1608DO_2OP_ACC_SCALAR_U(vmla, DO_VMLA)
   1609
   1610/* Vector by vector plus scalar */
   1611#define DO_VMLAS(D, N, M) ((N) * (D) + (M))
   1612
   1613DO_2OP_ACC_SCALAR_U(vmlas, DO_VMLAS)
   1614
   1615/*
   1616 * Long saturating scalar ops. As with DO_2OP_L, TYPE and H are for the
   1617 * input (smaller) type and LESIZE, LTYPE, LH for the output (long) type.
   1618 * SATMASK specifies which bits of the predicate mask matter for determining
   1619 * whether to propagate a saturation indication into FPSCR.QC -- for
   1620 * the 16x16->32 case we must check only the bit corresponding to the T or B
   1621 * half that we used, but for the 32x32->64 case we propagate if the mask
   1622 * bit is set for either half.
   1623 */
   1624#define DO_2OP_SAT_SCALAR_L(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN, SATMASK) \
   1625    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1626                                uint32_t rm)                            \
   1627    {                                                                   \
   1628        LTYPE *d = vd;                                                  \
   1629        TYPE *n = vn;                                                   \
   1630        TYPE m = rm;                                                    \
   1631        uint16_t mask = mve_element_mask(env);                          \
   1632        unsigned le;                                                    \
   1633        bool qc = false;                                                \
   1634        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
   1635            bool sat = false;                                           \
   1636            LTYPE r = FN((LTYPE)n[H##ESIZE(le * 2 + TOP)], m, &sat);    \
   1637            mergemask(&d[H##LESIZE(le)], r, mask);                      \
   1638            qc |= sat && (mask & SATMASK);                              \
   1639        }                                                               \
   1640        if (qc) {                                                       \
   1641            env->vfp.qc[0] = qc;                                        \
   1642        }                                                               \
   1643        mve_advance_vpt(env);                                           \
   1644    }
   1645
   1646static inline int32_t do_qdmullh(int16_t n, int16_t m, bool *sat)
   1647{
   1648    int64_t r = ((int64_t)n * m) * 2;
   1649    return do_sat_bhw(r, INT32_MIN, INT32_MAX, sat);
   1650}
   1651
   1652static inline int64_t do_qdmullw(int32_t n, int32_t m, bool *sat)
   1653{
   1654    /* The multiply can't overflow, but the doubling might */
   1655    int64_t r = (int64_t)n * m;
   1656    if (r > INT64_MAX / 2) {
   1657        *sat = true;
   1658        return INT64_MAX;
   1659    } else if (r < INT64_MIN / 2) {
   1660        *sat = true;
   1661        return INT64_MIN;
   1662    } else {
   1663        return r * 2;
   1664    }
   1665}
   1666
   1667#define SATMASK16B 1
   1668#define SATMASK16T (1 << 2)
   1669#define SATMASK32 ((1 << 4) | 1)
   1670
   1671DO_2OP_SAT_SCALAR_L(vqdmullb_scalarh, 0, 2, int16_t, 4, int32_t, \
   1672                    do_qdmullh, SATMASK16B)
   1673DO_2OP_SAT_SCALAR_L(vqdmullb_scalarw, 0, 4, int32_t, 8, int64_t, \
   1674                    do_qdmullw, SATMASK32)
   1675DO_2OP_SAT_SCALAR_L(vqdmullt_scalarh, 1, 2, int16_t, 4, int32_t, \
   1676                    do_qdmullh, SATMASK16T)
   1677DO_2OP_SAT_SCALAR_L(vqdmullt_scalarw, 1, 4, int32_t, 8, int64_t, \
   1678                    do_qdmullw, SATMASK32)
   1679
   1680/*
   1681 * Long saturating ops
   1682 */
   1683#define DO_2OP_SAT_L(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN, SATMASK)  \
   1684    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
   1685                                void *vm)                               \
   1686    {                                                                   \
   1687        LTYPE *d = vd;                                                  \
   1688        TYPE *n = vn, *m = vm;                                          \
   1689        uint16_t mask = mve_element_mask(env);                          \
   1690        unsigned le;                                                    \
   1691        bool qc = false;                                                \
   1692        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
   1693            bool sat = false;                                           \
   1694            LTYPE op1 = n[H##ESIZE(le * 2 + TOP)];                      \
   1695            LTYPE op2 = m[H##ESIZE(le * 2 + TOP)];                      \
   1696            mergemask(&d[H##LESIZE(le)], FN(op1, op2, &sat), mask);     \
   1697            qc |= sat && (mask & SATMASK);                              \
   1698        }                                                               \
   1699        if (qc) {                                                       \
   1700            env->vfp.qc[0] = qc;                                        \
   1701        }                                                               \
   1702        mve_advance_vpt(env);                                           \
   1703    }
   1704
   1705DO_2OP_SAT_L(vqdmullbh, 0, 2, int16_t, 4, int32_t, do_qdmullh, SATMASK16B)
   1706DO_2OP_SAT_L(vqdmullbw, 0, 4, int32_t, 8, int64_t, do_qdmullw, SATMASK32)
   1707DO_2OP_SAT_L(vqdmullth, 1, 2, int16_t, 4, int32_t, do_qdmullh, SATMASK16T)
   1708DO_2OP_SAT_L(vqdmulltw, 1, 4, int32_t, 8, int64_t, do_qdmullw, SATMASK32)
   1709
   1710static inline uint32_t do_vbrsrb(uint32_t n, uint32_t m)
   1711{
   1712    m &= 0xff;
   1713    if (m == 0) {
   1714        return 0;
   1715    }
   1716    n = revbit8(n);
   1717    if (m < 8) {
   1718        n >>= 8 - m;
   1719    }
   1720    return n;
   1721}
   1722
   1723static inline uint32_t do_vbrsrh(uint32_t n, uint32_t m)
   1724{
   1725    m &= 0xff;
   1726    if (m == 0) {
   1727        return 0;
   1728    }
   1729    n = revbit16(n);
   1730    if (m < 16) {
   1731        n >>= 16 - m;
   1732    }
   1733    return n;
   1734}
   1735
   1736static inline uint32_t do_vbrsrw(uint32_t n, uint32_t m)
   1737{
   1738    m &= 0xff;
   1739    if (m == 0) {
   1740        return 0;
   1741    }
   1742    n = revbit32(n);
   1743    if (m < 32) {
   1744        n >>= 32 - m;
   1745    }
   1746    return n;
   1747}
   1748
   1749DO_2OP_SCALAR(vbrsrb, 1, uint8_t, do_vbrsrb)
   1750DO_2OP_SCALAR(vbrsrh, 2, uint16_t, do_vbrsrh)
   1751DO_2OP_SCALAR(vbrsrw, 4, uint32_t, do_vbrsrw)
   1752
   1753/*
   1754 * Multiply add long dual accumulate ops.
   1755 */
   1756#define DO_LDAV(OP, ESIZE, TYPE, XCHG, EVENACC, ODDACC)                 \
   1757    uint64_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vn,         \
   1758                                    void *vm, uint64_t a)               \
   1759    {                                                                   \
   1760        uint16_t mask = mve_element_mask(env);                          \
   1761        unsigned e;                                                     \
   1762        TYPE *n = vn, *m = vm;                                          \
   1763        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1764            if (mask & 1) {                                             \
   1765                if (e & 1) {                                            \
   1766                    a ODDACC                                            \
   1767                        (int64_t)n[H##ESIZE(e - 1 * XCHG)] * m[H##ESIZE(e)]; \
   1768                } else {                                                \
   1769                    a EVENACC                                           \
   1770                        (int64_t)n[H##ESIZE(e + 1 * XCHG)] * m[H##ESIZE(e)]; \
   1771                }                                                       \
   1772            }                                                           \
   1773        }                                                               \
   1774        mve_advance_vpt(env);                                           \
   1775        return a;                                                       \
   1776    }
   1777
   1778DO_LDAV(vmlaldavsh, 2, int16_t, false, +=, +=)
   1779DO_LDAV(vmlaldavxsh, 2, int16_t, true, +=, +=)
   1780DO_LDAV(vmlaldavsw, 4, int32_t, false, +=, +=)
   1781DO_LDAV(vmlaldavxsw, 4, int32_t, true, +=, +=)
   1782
   1783DO_LDAV(vmlaldavuh, 2, uint16_t, false, +=, +=)
   1784DO_LDAV(vmlaldavuw, 4, uint32_t, false, +=, +=)
   1785
   1786DO_LDAV(vmlsldavsh, 2, int16_t, false, +=, -=)
   1787DO_LDAV(vmlsldavxsh, 2, int16_t, true, +=, -=)
   1788DO_LDAV(vmlsldavsw, 4, int32_t, false, +=, -=)
   1789DO_LDAV(vmlsldavxsw, 4, int32_t, true, +=, -=)
   1790
   1791/*
   1792 * Multiply add dual accumulate ops
   1793 */
   1794#define DO_DAV(OP, ESIZE, TYPE, XCHG, EVENACC, ODDACC) \
   1795    uint32_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vn,         \
   1796                                    void *vm, uint32_t a)               \
   1797    {                                                                   \
   1798        uint16_t mask = mve_element_mask(env);                          \
   1799        unsigned e;                                                     \
   1800        TYPE *n = vn, *m = vm;                                          \
   1801        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   1802            if (mask & 1) {                                             \
   1803                if (e & 1) {                                            \
   1804                    a ODDACC                                            \
   1805                        n[H##ESIZE(e - 1 * XCHG)] * m[H##ESIZE(e)];     \
   1806                } else {                                                \
   1807                    a EVENACC                                           \
   1808                        n[H##ESIZE(e + 1 * XCHG)] * m[H##ESIZE(e)];     \
   1809                }                                                       \
   1810            }                                                           \
   1811        }                                                               \
   1812        mve_advance_vpt(env);                                           \
   1813        return a;                                                       \
   1814    }
   1815
   1816#define DO_DAV_S(INSN, XCHG, EVENACC, ODDACC)           \
   1817    DO_DAV(INSN##b, 1, int8_t, XCHG, EVENACC, ODDACC)   \
   1818    DO_DAV(INSN##h, 2, int16_t, XCHG, EVENACC, ODDACC)  \
   1819    DO_DAV(INSN##w, 4, int32_t, XCHG, EVENACC, ODDACC)
   1820
   1821#define DO_DAV_U(INSN, XCHG, EVENACC, ODDACC)           \
   1822    DO_DAV(INSN##b, 1, uint8_t, XCHG, EVENACC, ODDACC)  \
   1823    DO_DAV(INSN##h, 2, uint16_t, XCHG, EVENACC, ODDACC) \
   1824    DO_DAV(INSN##w, 4, uint32_t, XCHG, EVENACC, ODDACC)
   1825
   1826DO_DAV_S(vmladavs, false, +=, +=)
   1827DO_DAV_U(vmladavu, false, +=, +=)
   1828DO_DAV_S(vmlsdav, false, +=, -=)
   1829DO_DAV_S(vmladavsx, true, +=, +=)
   1830DO_DAV_S(vmlsdavx, true, +=, -=)
   1831
   1832/*
   1833 * Rounding multiply add long dual accumulate high. In the pseudocode
   1834 * this is implemented with a 72-bit internal accumulator value of which
   1835 * the top 64 bits are returned. We optimize this to avoid having to
   1836 * use 128-bit arithmetic -- we can do this because the 74-bit accumulator
   1837 * is squashed back into 64-bits after each beat.
   1838 */
   1839#define DO_LDAVH(OP, TYPE, LTYPE, XCHG, SUB)                            \
   1840    uint64_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vn,         \
   1841                                    void *vm, uint64_t a)               \
   1842    {                                                                   \
   1843        uint16_t mask = mve_element_mask(env);                          \
   1844        unsigned e;                                                     \
   1845        TYPE *n = vn, *m = vm;                                          \
   1846        for (e = 0; e < 16 / 4; e++, mask >>= 4) {                      \
   1847            if (mask & 1) {                                             \
   1848                LTYPE mul;                                              \
   1849                if (e & 1) {                                            \
   1850                    mul = (LTYPE)n[H4(e - 1 * XCHG)] * m[H4(e)];        \
   1851                    if (SUB) {                                          \
   1852                        mul = -mul;                                     \
   1853                    }                                                   \
   1854                } else {                                                \
   1855                    mul = (LTYPE)n[H4(e + 1 * XCHG)] * m[H4(e)];        \
   1856                }                                                       \
   1857                mul = (mul >> 8) + ((mul >> 7) & 1);                    \
   1858                a += mul;                                               \
   1859            }                                                           \
   1860        }                                                               \
   1861        mve_advance_vpt(env);                                           \
   1862        return a;                                                       \
   1863    }
   1864
   1865DO_LDAVH(vrmlaldavhsw, int32_t, int64_t, false, false)
   1866DO_LDAVH(vrmlaldavhxsw, int32_t, int64_t, true, false)
   1867
   1868DO_LDAVH(vrmlaldavhuw, uint32_t, uint64_t, false, false)
   1869
   1870DO_LDAVH(vrmlsldavhsw, int32_t, int64_t, false, true)
   1871DO_LDAVH(vrmlsldavhxsw, int32_t, int64_t, true, true)
   1872
   1873/* Vector add across vector */
   1874#define DO_VADDV(OP, ESIZE, TYPE)                               \
   1875    uint32_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vm, \
   1876                                    uint32_t ra)                \
   1877    {                                                           \
   1878        uint16_t mask = mve_element_mask(env);                  \
   1879        unsigned e;                                             \
   1880        TYPE *m = vm;                                           \
   1881        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   1882            if (mask & 1) {                                     \
   1883                ra += m[H##ESIZE(e)];                           \
   1884            }                                                   \
   1885        }                                                       \
   1886        mve_advance_vpt(env);                                   \
   1887        return ra;                                              \
   1888    }                                                           \
   1889
   1890DO_VADDV(vaddvsb, 1, int8_t)
   1891DO_VADDV(vaddvsh, 2, int16_t)
   1892DO_VADDV(vaddvsw, 4, int32_t)
   1893DO_VADDV(vaddvub, 1, uint8_t)
   1894DO_VADDV(vaddvuh, 2, uint16_t)
   1895DO_VADDV(vaddvuw, 4, uint32_t)
   1896
   1897/*
   1898 * Vector max/min across vector. Unlike VADDV, we must
   1899 * read ra as the element size, not its full width.
   1900 * We work with int64_t internally for simplicity.
   1901 */
   1902#define DO_VMAXMINV(OP, ESIZE, TYPE, RATYPE, FN)                \
   1903    uint32_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vm, \
   1904                                    uint32_t ra_in)             \
   1905    {                                                           \
   1906        uint16_t mask = mve_element_mask(env);                  \
   1907        unsigned e;                                             \
   1908        TYPE *m = vm;                                           \
   1909        int64_t ra = (RATYPE)ra_in;                             \
   1910        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   1911            if (mask & 1) {                                     \
   1912                ra = FN(ra, m[H##ESIZE(e)]);                    \
   1913            }                                                   \
   1914        }                                                       \
   1915        mve_advance_vpt(env);                                   \
   1916        return ra;                                              \
   1917    }                                                           \
   1918
   1919#define DO_VMAXMINV_U(INSN, FN)                         \
   1920    DO_VMAXMINV(INSN##b, 1, uint8_t, uint8_t, FN)       \
   1921    DO_VMAXMINV(INSN##h, 2, uint16_t, uint16_t, FN)     \
   1922    DO_VMAXMINV(INSN##w, 4, uint32_t, uint32_t, FN)
   1923#define DO_VMAXMINV_S(INSN, FN)                         \
   1924    DO_VMAXMINV(INSN##b, 1, int8_t, int8_t, FN)         \
   1925    DO_VMAXMINV(INSN##h, 2, int16_t, int16_t, FN)       \
   1926    DO_VMAXMINV(INSN##w, 4, int32_t, int32_t, FN)
   1927
   1928/*
   1929 * Helpers for max and min of absolute values across vector:
   1930 * note that we only take the absolute value of 'm', not 'n'
   1931 */
   1932static int64_t do_maxa(int64_t n, int64_t m)
   1933{
   1934    if (m < 0) {
   1935        m = -m;
   1936    }
   1937    return MAX(n, m);
   1938}
   1939
   1940static int64_t do_mina(int64_t n, int64_t m)
   1941{
   1942    if (m < 0) {
   1943        m = -m;
   1944    }
   1945    return MIN(n, m);
   1946}
   1947
   1948DO_VMAXMINV_S(vmaxvs, DO_MAX)
   1949DO_VMAXMINV_U(vmaxvu, DO_MAX)
   1950DO_VMAXMINV_S(vminvs, DO_MIN)
   1951DO_VMAXMINV_U(vminvu, DO_MIN)
   1952/*
   1953 * VMAXAV, VMINAV treat the general purpose input as unsigned
   1954 * and the vector elements as signed.
   1955 */
   1956DO_VMAXMINV(vmaxavb, 1, int8_t, uint8_t, do_maxa)
   1957DO_VMAXMINV(vmaxavh, 2, int16_t, uint16_t, do_maxa)
   1958DO_VMAXMINV(vmaxavw, 4, int32_t, uint32_t, do_maxa)
   1959DO_VMAXMINV(vminavb, 1, int8_t, uint8_t, do_mina)
   1960DO_VMAXMINV(vminavh, 2, int16_t, uint16_t, do_mina)
   1961DO_VMAXMINV(vminavw, 4, int32_t, uint32_t, do_mina)
   1962
   1963#define DO_VABAV(OP, ESIZE, TYPE)                               \
   1964    uint32_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vn, \
   1965                                    void *vm, uint32_t ra)      \
   1966    {                                                           \
   1967        uint16_t mask = mve_element_mask(env);                  \
   1968        unsigned e;                                             \
   1969        TYPE *m = vm, *n = vn;                                  \
   1970        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   1971            if (mask & 1) {                                     \
   1972                int64_t n0 = n[H##ESIZE(e)];                    \
   1973                int64_t m0 = m[H##ESIZE(e)];                    \
   1974                uint32_t r = n0 >= m0 ? (n0 - m0) : (m0 - n0);  \
   1975                ra += r;                                        \
   1976            }                                                   \
   1977        }                                                       \
   1978        mve_advance_vpt(env);                                   \
   1979        return ra;                                              \
   1980    }
   1981
   1982DO_VABAV(vabavsb, 1, int8_t)
   1983DO_VABAV(vabavsh, 2, int16_t)
   1984DO_VABAV(vabavsw, 4, int32_t)
   1985DO_VABAV(vabavub, 1, uint8_t)
   1986DO_VABAV(vabavuh, 2, uint16_t)
   1987DO_VABAV(vabavuw, 4, uint32_t)
   1988
   1989#define DO_VADDLV(OP, TYPE, LTYPE)                              \
   1990    uint64_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vm, \
   1991                                    uint64_t ra)                \
   1992    {                                                           \
   1993        uint16_t mask = mve_element_mask(env);                  \
   1994        unsigned e;                                             \
   1995        TYPE *m = vm;                                           \
   1996        for (e = 0; e < 16 / 4; e++, mask >>= 4) {              \
   1997            if (mask & 1) {                                     \
   1998                ra += (LTYPE)m[H4(e)];                          \
   1999            }                                                   \
   2000        }                                                       \
   2001        mve_advance_vpt(env);                                   \
   2002        return ra;                                              \
   2003    }                                                           \
   2004
   2005DO_VADDLV(vaddlv_s, int32_t, int64_t)
   2006DO_VADDLV(vaddlv_u, uint32_t, uint64_t)
   2007
   2008/* Shifts by immediate */
   2009#define DO_2SHIFT(OP, ESIZE, TYPE, FN)                          \
   2010    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,     \
   2011                                void *vm, uint32_t shift)       \
   2012    {                                                           \
   2013        TYPE *d = vd, *m = vm;                                  \
   2014        uint16_t mask = mve_element_mask(env);                  \
   2015        unsigned e;                                             \
   2016        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   2017            mergemask(&d[H##ESIZE(e)],                          \
   2018                      FN(m[H##ESIZE(e)], shift), mask);         \
   2019        }                                                       \
   2020        mve_advance_vpt(env);                                   \
   2021    }
   2022
   2023#define DO_2SHIFT_SAT(OP, ESIZE, TYPE, FN)                      \
   2024    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,     \
   2025                                void *vm, uint32_t shift)       \
   2026    {                                                           \
   2027        TYPE *d = vd, *m = vm;                                  \
   2028        uint16_t mask = mve_element_mask(env);                  \
   2029        unsigned e;                                             \
   2030        bool qc = false;                                        \
   2031        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   2032            bool sat = false;                                   \
   2033            mergemask(&d[H##ESIZE(e)],                          \
   2034                      FN(m[H##ESIZE(e)], shift, &sat), mask);   \
   2035            qc |= sat & mask & 1;                               \
   2036        }                                                       \
   2037        if (qc) {                                               \
   2038            env->vfp.qc[0] = qc;                                \
   2039        }                                                       \
   2040        mve_advance_vpt(env);                                   \
   2041    }
   2042
   2043/* provide unsigned 2-op shift helpers for all sizes */
   2044#define DO_2SHIFT_U(OP, FN)                     \
   2045    DO_2SHIFT(OP##b, 1, uint8_t, FN)            \
   2046    DO_2SHIFT(OP##h, 2, uint16_t, FN)           \
   2047    DO_2SHIFT(OP##w, 4, uint32_t, FN)
   2048#define DO_2SHIFT_S(OP, FN)                     \
   2049    DO_2SHIFT(OP##b, 1, int8_t, FN)             \
   2050    DO_2SHIFT(OP##h, 2, int16_t, FN)            \
   2051    DO_2SHIFT(OP##w, 4, int32_t, FN)
   2052
   2053#define DO_2SHIFT_SAT_U(OP, FN)                 \
   2054    DO_2SHIFT_SAT(OP##b, 1, uint8_t, FN)        \
   2055    DO_2SHIFT_SAT(OP##h, 2, uint16_t, FN)       \
   2056    DO_2SHIFT_SAT(OP##w, 4, uint32_t, FN)
   2057#define DO_2SHIFT_SAT_S(OP, FN)                 \
   2058    DO_2SHIFT_SAT(OP##b, 1, int8_t, FN)         \
   2059    DO_2SHIFT_SAT(OP##h, 2, int16_t, FN)        \
   2060    DO_2SHIFT_SAT(OP##w, 4, int32_t, FN)
   2061
   2062DO_2SHIFT_U(vshli_u, DO_VSHLU)
   2063DO_2SHIFT_S(vshli_s, DO_VSHLS)
   2064DO_2SHIFT_SAT_U(vqshli_u, DO_UQSHL_OP)
   2065DO_2SHIFT_SAT_S(vqshli_s, DO_SQSHL_OP)
   2066DO_2SHIFT_SAT_S(vqshlui_s, DO_SUQSHL_OP)
   2067DO_2SHIFT_U(vrshli_u, DO_VRSHLU)
   2068DO_2SHIFT_S(vrshli_s, DO_VRSHLS)
   2069DO_2SHIFT_SAT_U(vqrshli_u, DO_UQRSHL_OP)
   2070DO_2SHIFT_SAT_S(vqrshli_s, DO_SQRSHL_OP)
   2071
   2072/* Shift-and-insert; we always work with 64 bits at a time */
   2073#define DO_2SHIFT_INSERT(OP, ESIZE, SHIFTFN, MASKFN)                    \
   2074    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,             \
   2075                                void *vm, uint32_t shift)               \
   2076    {                                                                   \
   2077        uint64_t *d = vd, *m = vm;                                      \
   2078        uint16_t mask;                                                  \
   2079        uint64_t shiftmask;                                             \
   2080        unsigned e;                                                     \
   2081        if (shift == ESIZE * 8) {                                       \
   2082            /*                                                          \
   2083             * Only VSRI can shift by <dt>; it should mean "don't       \
   2084             * update the destination". The generic logic can't handle  \
   2085             * this because it would try to shift by an out-of-range    \
   2086             * amount, so special case it here.                         \
   2087             */                                                         \
   2088            goto done;                                                  \
   2089        }                                                               \
   2090        assert(shift < ESIZE * 8);                                      \
   2091        mask = mve_element_mask(env);                                   \
   2092        /* ESIZE / 2 gives the MO_* value if ESIZE is in [1,2,4] */     \
   2093        shiftmask = dup_const(ESIZE / 2, MASKFN(ESIZE * 8, shift));     \
   2094        for (e = 0; e < 16 / 8; e++, mask >>= 8) {                      \
   2095            uint64_t r = (SHIFTFN(m[H8(e)], shift) & shiftmask) |       \
   2096                (d[H8(e)] & ~shiftmask);                                \
   2097            mergemask(&d[H8(e)], r, mask);                              \
   2098        }                                                               \
   2099done:                                                                   \
   2100        mve_advance_vpt(env);                                           \
   2101    }
   2102
   2103#define DO_SHL(N, SHIFT) ((N) << (SHIFT))
   2104#define DO_SHR(N, SHIFT) ((N) >> (SHIFT))
   2105#define SHL_MASK(EBITS, SHIFT) MAKE_64BIT_MASK((SHIFT), (EBITS) - (SHIFT))
   2106#define SHR_MASK(EBITS, SHIFT) MAKE_64BIT_MASK(0, (EBITS) - (SHIFT))
   2107
   2108DO_2SHIFT_INSERT(vsrib, 1, DO_SHR, SHR_MASK)
   2109DO_2SHIFT_INSERT(vsrih, 2, DO_SHR, SHR_MASK)
   2110DO_2SHIFT_INSERT(vsriw, 4, DO_SHR, SHR_MASK)
   2111DO_2SHIFT_INSERT(vslib, 1, DO_SHL, SHL_MASK)
   2112DO_2SHIFT_INSERT(vslih, 2, DO_SHL, SHL_MASK)
   2113DO_2SHIFT_INSERT(vsliw, 4, DO_SHL, SHL_MASK)
   2114
   2115/*
   2116 * Long shifts taking half-sized inputs from top or bottom of the input
   2117 * vector and producing a double-width result. ESIZE, TYPE are for
   2118 * the input, and LESIZE, LTYPE for the output.
   2119 * Unlike the normal shift helpers, we do not handle negative shift counts,
   2120 * because the long shift is strictly left-only.
   2121 */
   2122#define DO_VSHLL(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE)                   \
   2123    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,             \
   2124                                void *vm, uint32_t shift)               \
   2125    {                                                                   \
   2126        LTYPE *d = vd;                                                  \
   2127        TYPE *m = vm;                                                   \
   2128        uint16_t mask = mve_element_mask(env);                          \
   2129        unsigned le;                                                    \
   2130        assert(shift <= 16);                                            \
   2131        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
   2132            LTYPE r = (LTYPE)m[H##ESIZE(le * 2 + TOP)] << shift;        \
   2133            mergemask(&d[H##LESIZE(le)], r, mask);                      \
   2134        }                                                               \
   2135        mve_advance_vpt(env);                                           \
   2136    }
   2137
   2138#define DO_VSHLL_ALL(OP, TOP)                                \
   2139    DO_VSHLL(OP##sb, TOP, 1, int8_t, 2, int16_t)             \
   2140    DO_VSHLL(OP##ub, TOP, 1, uint8_t, 2, uint16_t)           \
   2141    DO_VSHLL(OP##sh, TOP, 2, int16_t, 4, int32_t)            \
   2142    DO_VSHLL(OP##uh, TOP, 2, uint16_t, 4, uint32_t)          \
   2143
   2144DO_VSHLL_ALL(vshllb, false)
   2145DO_VSHLL_ALL(vshllt, true)
   2146
   2147/*
   2148 * Narrowing right shifts, taking a double sized input, shifting it
   2149 * and putting the result in either the top or bottom half of the output.
   2150 * ESIZE, TYPE are the output, and LESIZE, LTYPE the input.
   2151 */
   2152#define DO_VSHRN(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN)       \
   2153    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,     \
   2154                                void *vm, uint32_t shift)       \
   2155    {                                                           \
   2156        LTYPE *m = vm;                                          \
   2157        TYPE *d = vd;                                           \
   2158        uint16_t mask = mve_element_mask(env);                  \
   2159        unsigned le;                                            \
   2160        mask >>= ESIZE * TOP;                                   \
   2161        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) { \
   2162            TYPE r = FN(m[H##LESIZE(le)], shift);               \
   2163            mergemask(&d[H##ESIZE(le * 2 + TOP)], r, mask);     \
   2164        }                                                       \
   2165        mve_advance_vpt(env);                                   \
   2166    }
   2167
   2168#define DO_VSHRN_ALL(OP, FN)                                    \
   2169    DO_VSHRN(OP##bb, false, 1, uint8_t, 2, uint16_t, FN)        \
   2170    DO_VSHRN(OP##bh, false, 2, uint16_t, 4, uint32_t, FN)       \
   2171    DO_VSHRN(OP##tb, true, 1, uint8_t, 2, uint16_t, FN)         \
   2172    DO_VSHRN(OP##th, true, 2, uint16_t, 4, uint32_t, FN)
   2173
   2174static inline uint64_t do_urshr(uint64_t x, unsigned sh)
   2175{
   2176    if (likely(sh < 64)) {
   2177        return (x >> sh) + ((x >> (sh - 1)) & 1);
   2178    } else if (sh == 64) {
   2179        return x >> 63;
   2180    } else {
   2181        return 0;
   2182    }
   2183}
   2184
   2185static inline int64_t do_srshr(int64_t x, unsigned sh)
   2186{
   2187    if (likely(sh < 64)) {
   2188        return (x >> sh) + ((x >> (sh - 1)) & 1);
   2189    } else {
   2190        /* Rounding the sign bit always produces 0. */
   2191        return 0;
   2192    }
   2193}
   2194
   2195DO_VSHRN_ALL(vshrn, DO_SHR)
   2196DO_VSHRN_ALL(vrshrn, do_urshr)
   2197
   2198static inline int32_t do_sat_bhs(int64_t val, int64_t min, int64_t max,
   2199                                 bool *satp)
   2200{
   2201    if (val > max) {
   2202        *satp = true;
   2203        return max;
   2204    } else if (val < min) {
   2205        *satp = true;
   2206        return min;
   2207    } else {
   2208        return val;
   2209    }
   2210}
   2211
   2212/* Saturating narrowing right shifts */
   2213#define DO_VSHRN_SAT(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN)   \
   2214    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd,     \
   2215                                void *vm, uint32_t shift)       \
   2216    {                                                           \
   2217        LTYPE *m = vm;                                          \
   2218        TYPE *d = vd;                                           \
   2219        uint16_t mask = mve_element_mask(env);                  \
   2220        bool qc = false;                                        \
   2221        unsigned le;                                            \
   2222        mask >>= ESIZE * TOP;                                   \
   2223        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) { \
   2224            bool sat = false;                                   \
   2225            TYPE r = FN(m[H##LESIZE(le)], shift, &sat);         \
   2226            mergemask(&d[H##ESIZE(le * 2 + TOP)], r, mask);     \
   2227            qc |= sat & mask & 1;                               \
   2228        }                                                       \
   2229        if (qc) {                                               \
   2230            env->vfp.qc[0] = qc;                                \
   2231        }                                                       \
   2232        mve_advance_vpt(env);                                   \
   2233    }
   2234
   2235#define DO_VSHRN_SAT_UB(BOP, TOP, FN)                           \
   2236    DO_VSHRN_SAT(BOP, false, 1, uint8_t, 2, uint16_t, FN)       \
   2237    DO_VSHRN_SAT(TOP, true, 1, uint8_t, 2, uint16_t, FN)
   2238
   2239#define DO_VSHRN_SAT_UH(BOP, TOP, FN)                           \
   2240    DO_VSHRN_SAT(BOP, false, 2, uint16_t, 4, uint32_t, FN)      \
   2241    DO_VSHRN_SAT(TOP, true, 2, uint16_t, 4, uint32_t, FN)
   2242
   2243#define DO_VSHRN_SAT_SB(BOP, TOP, FN)                           \
   2244    DO_VSHRN_SAT(BOP, false, 1, int8_t, 2, int16_t, FN)         \
   2245    DO_VSHRN_SAT(TOP, true, 1, int8_t, 2, int16_t, FN)
   2246
   2247#define DO_VSHRN_SAT_SH(BOP, TOP, FN)                           \
   2248    DO_VSHRN_SAT(BOP, false, 2, int16_t, 4, int32_t, FN)        \
   2249    DO_VSHRN_SAT(TOP, true, 2, int16_t, 4, int32_t, FN)
   2250
   2251#define DO_SHRN_SB(N, M, SATP)                                  \
   2252    do_sat_bhs((int64_t)(N) >> (M), INT8_MIN, INT8_MAX, SATP)
   2253#define DO_SHRN_UB(N, M, SATP)                                  \
   2254    do_sat_bhs((uint64_t)(N) >> (M), 0, UINT8_MAX, SATP)
   2255#define DO_SHRUN_B(N, M, SATP)                                  \
   2256    do_sat_bhs((int64_t)(N) >> (M), 0, UINT8_MAX, SATP)
   2257
   2258#define DO_SHRN_SH(N, M, SATP)                                  \
   2259    do_sat_bhs((int64_t)(N) >> (M), INT16_MIN, INT16_MAX, SATP)
   2260#define DO_SHRN_UH(N, M, SATP)                                  \
   2261    do_sat_bhs((uint64_t)(N) >> (M), 0, UINT16_MAX, SATP)
   2262#define DO_SHRUN_H(N, M, SATP)                                  \
   2263    do_sat_bhs((int64_t)(N) >> (M), 0, UINT16_MAX, SATP)
   2264
   2265#define DO_RSHRN_SB(N, M, SATP)                                 \
   2266    do_sat_bhs(do_srshr(N, M), INT8_MIN, INT8_MAX, SATP)
   2267#define DO_RSHRN_UB(N, M, SATP)                                 \
   2268    do_sat_bhs(do_urshr(N, M), 0, UINT8_MAX, SATP)
   2269#define DO_RSHRUN_B(N, M, SATP)                                 \
   2270    do_sat_bhs(do_srshr(N, M), 0, UINT8_MAX, SATP)
   2271
   2272#define DO_RSHRN_SH(N, M, SATP)                                 \
   2273    do_sat_bhs(do_srshr(N, M), INT16_MIN, INT16_MAX, SATP)
   2274#define DO_RSHRN_UH(N, M, SATP)                                 \
   2275    do_sat_bhs(do_urshr(N, M), 0, UINT16_MAX, SATP)
   2276#define DO_RSHRUN_H(N, M, SATP)                                 \
   2277    do_sat_bhs(do_srshr(N, M), 0, UINT16_MAX, SATP)
   2278
   2279DO_VSHRN_SAT_SB(vqshrnb_sb, vqshrnt_sb, DO_SHRN_SB)
   2280DO_VSHRN_SAT_SH(vqshrnb_sh, vqshrnt_sh, DO_SHRN_SH)
   2281DO_VSHRN_SAT_UB(vqshrnb_ub, vqshrnt_ub, DO_SHRN_UB)
   2282DO_VSHRN_SAT_UH(vqshrnb_uh, vqshrnt_uh, DO_SHRN_UH)
   2283DO_VSHRN_SAT_SB(vqshrunbb, vqshruntb, DO_SHRUN_B)
   2284DO_VSHRN_SAT_SH(vqshrunbh, vqshrunth, DO_SHRUN_H)
   2285
   2286DO_VSHRN_SAT_SB(vqrshrnb_sb, vqrshrnt_sb, DO_RSHRN_SB)
   2287DO_VSHRN_SAT_SH(vqrshrnb_sh, vqrshrnt_sh, DO_RSHRN_SH)
   2288DO_VSHRN_SAT_UB(vqrshrnb_ub, vqrshrnt_ub, DO_RSHRN_UB)
   2289DO_VSHRN_SAT_UH(vqrshrnb_uh, vqrshrnt_uh, DO_RSHRN_UH)
   2290DO_VSHRN_SAT_SB(vqrshrunbb, vqrshruntb, DO_RSHRUN_B)
   2291DO_VSHRN_SAT_SH(vqrshrunbh, vqrshrunth, DO_RSHRUN_H)
   2292
   2293#define DO_VMOVN(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE)                   \
   2294    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm)         \
   2295    {                                                                   \
   2296        LTYPE *m = vm;                                                  \
   2297        TYPE *d = vd;                                                   \
   2298        uint16_t mask = mve_element_mask(env);                          \
   2299        unsigned le;                                                    \
   2300        mask >>= ESIZE * TOP;                                           \
   2301        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
   2302            mergemask(&d[H##ESIZE(le * 2 + TOP)],                       \
   2303                      m[H##LESIZE(le)], mask);                          \
   2304        }                                                               \
   2305        mve_advance_vpt(env);                                           \
   2306    }
   2307
   2308DO_VMOVN(vmovnbb, false, 1, uint8_t, 2, uint16_t)
   2309DO_VMOVN(vmovnbh, false, 2, uint16_t, 4, uint32_t)
   2310DO_VMOVN(vmovntb, true, 1, uint8_t, 2, uint16_t)
   2311DO_VMOVN(vmovnth, true, 2, uint16_t, 4, uint32_t)
   2312
   2313#define DO_VMOVN_SAT(OP, TOP, ESIZE, TYPE, LESIZE, LTYPE, FN)           \
   2314    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm)         \
   2315    {                                                                   \
   2316        LTYPE *m = vm;                                                  \
   2317        TYPE *d = vd;                                                   \
   2318        uint16_t mask = mve_element_mask(env);                          \
   2319        bool qc = false;                                                \
   2320        unsigned le;                                                    \
   2321        mask >>= ESIZE * TOP;                                           \
   2322        for (le = 0; le < 16 / LESIZE; le++, mask >>= LESIZE) {         \
   2323            bool sat = false;                                           \
   2324            TYPE r = FN(m[H##LESIZE(le)], &sat);                        \
   2325            mergemask(&d[H##ESIZE(le * 2 + TOP)], r, mask);             \
   2326            qc |= sat & mask & 1;                                       \
   2327        }                                                               \
   2328        if (qc) {                                                       \
   2329            env->vfp.qc[0] = qc;                                        \
   2330        }                                                               \
   2331        mve_advance_vpt(env);                                           \
   2332    }
   2333
   2334#define DO_VMOVN_SAT_UB(BOP, TOP, FN)                           \
   2335    DO_VMOVN_SAT(BOP, false, 1, uint8_t, 2, uint16_t, FN)       \
   2336    DO_VMOVN_SAT(TOP, true, 1, uint8_t, 2, uint16_t, FN)
   2337
   2338#define DO_VMOVN_SAT_UH(BOP, TOP, FN)                           \
   2339    DO_VMOVN_SAT(BOP, false, 2, uint16_t, 4, uint32_t, FN)      \
   2340    DO_VMOVN_SAT(TOP, true, 2, uint16_t, 4, uint32_t, FN)
   2341
   2342#define DO_VMOVN_SAT_SB(BOP, TOP, FN)                           \
   2343    DO_VMOVN_SAT(BOP, false, 1, int8_t, 2, int16_t, FN)         \
   2344    DO_VMOVN_SAT(TOP, true, 1, int8_t, 2, int16_t, FN)
   2345
   2346#define DO_VMOVN_SAT_SH(BOP, TOP, FN)                           \
   2347    DO_VMOVN_SAT(BOP, false, 2, int16_t, 4, int32_t, FN)        \
   2348    DO_VMOVN_SAT(TOP, true, 2, int16_t, 4, int32_t, FN)
   2349
   2350#define DO_VQMOVN_SB(N, SATP)                           \
   2351    do_sat_bhs((int64_t)(N), INT8_MIN, INT8_MAX, SATP)
   2352#define DO_VQMOVN_UB(N, SATP)                           \
   2353    do_sat_bhs((uint64_t)(N), 0, UINT8_MAX, SATP)
   2354#define DO_VQMOVUN_B(N, SATP)                           \
   2355    do_sat_bhs((int64_t)(N), 0, UINT8_MAX, SATP)
   2356
   2357#define DO_VQMOVN_SH(N, SATP)                           \
   2358    do_sat_bhs((int64_t)(N), INT16_MIN, INT16_MAX, SATP)
   2359#define DO_VQMOVN_UH(N, SATP)                           \
   2360    do_sat_bhs((uint64_t)(N), 0, UINT16_MAX, SATP)
   2361#define DO_VQMOVUN_H(N, SATP)                           \
   2362    do_sat_bhs((int64_t)(N), 0, UINT16_MAX, SATP)
   2363
   2364DO_VMOVN_SAT_SB(vqmovnbsb, vqmovntsb, DO_VQMOVN_SB)
   2365DO_VMOVN_SAT_SH(vqmovnbsh, vqmovntsh, DO_VQMOVN_SH)
   2366DO_VMOVN_SAT_UB(vqmovnbub, vqmovntub, DO_VQMOVN_UB)
   2367DO_VMOVN_SAT_UH(vqmovnbuh, vqmovntuh, DO_VQMOVN_UH)
   2368DO_VMOVN_SAT_SB(vqmovunbb, vqmovuntb, DO_VQMOVUN_B)
   2369DO_VMOVN_SAT_SH(vqmovunbh, vqmovunth, DO_VQMOVUN_H)
   2370
   2371uint32_t HELPER(mve_vshlc)(CPUARMState *env, void *vd, uint32_t rdm,
   2372                           uint32_t shift)
   2373{
   2374    uint32_t *d = vd;
   2375    uint16_t mask = mve_element_mask(env);
   2376    unsigned e;
   2377    uint32_t r;
   2378
   2379    /*
   2380     * For each 32-bit element, we shift it left, bringing in the
   2381     * low 'shift' bits of rdm at the bottom. Bits shifted out at
   2382     * the top become the new rdm, if the predicate mask permits.
   2383     * The final rdm value is returned to update the register.
   2384     * shift == 0 here means "shift by 32 bits".
   2385     */
   2386    if (shift == 0) {
   2387        for (e = 0; e < 16 / 4; e++, mask >>= 4) {
   2388            r = rdm;
   2389            if (mask & 1) {
   2390                rdm = d[H4(e)];
   2391            }
   2392            mergemask(&d[H4(e)], r, mask);
   2393        }
   2394    } else {
   2395        uint32_t shiftmask = MAKE_64BIT_MASK(0, shift);
   2396
   2397        for (e = 0; e < 16 / 4; e++, mask >>= 4) {
   2398            r = (d[H4(e)] << shift) | (rdm & shiftmask);
   2399            if (mask & 1) {
   2400                rdm = d[H4(e)] >> (32 - shift);
   2401            }
   2402            mergemask(&d[H4(e)], r, mask);
   2403        }
   2404    }
   2405    mve_advance_vpt(env);
   2406    return rdm;
   2407}
   2408
   2409uint64_t HELPER(mve_sshrl)(CPUARMState *env, uint64_t n, uint32_t shift)
   2410{
   2411    return do_sqrshl_d(n, -(int8_t)shift, false, NULL);
   2412}
   2413
   2414uint64_t HELPER(mve_ushll)(CPUARMState *env, uint64_t n, uint32_t shift)
   2415{
   2416    return do_uqrshl_d(n, (int8_t)shift, false, NULL);
   2417}
   2418
   2419uint64_t HELPER(mve_sqshll)(CPUARMState *env, uint64_t n, uint32_t shift)
   2420{
   2421    return do_sqrshl_d(n, (int8_t)shift, false, &env->QF);
   2422}
   2423
   2424uint64_t HELPER(mve_uqshll)(CPUARMState *env, uint64_t n, uint32_t shift)
   2425{
   2426    return do_uqrshl_d(n, (int8_t)shift, false, &env->QF);
   2427}
   2428
   2429uint64_t HELPER(mve_sqrshrl)(CPUARMState *env, uint64_t n, uint32_t shift)
   2430{
   2431    return do_sqrshl_d(n, -(int8_t)shift, true, &env->QF);
   2432}
   2433
   2434uint64_t HELPER(mve_uqrshll)(CPUARMState *env, uint64_t n, uint32_t shift)
   2435{
   2436    return do_uqrshl_d(n, (int8_t)shift, true, &env->QF);
   2437}
   2438
   2439/* Operate on 64-bit values, but saturate at 48 bits */
   2440static inline int64_t do_sqrshl48_d(int64_t src, int64_t shift,
   2441                                    bool round, uint32_t *sat)
   2442{
   2443    int64_t val, extval;
   2444
   2445    if (shift <= -48) {
   2446        /* Rounding the sign bit always produces 0. */
   2447        if (round) {
   2448            return 0;
   2449        }
   2450        return src >> 63;
   2451    } else if (shift < 0) {
   2452        if (round) {
   2453            src >>= -shift - 1;
   2454            val = (src >> 1) + (src & 1);
   2455        } else {
   2456            val = src >> -shift;
   2457        }
   2458        extval = sextract64(val, 0, 48);
   2459        if (!sat || val == extval) {
   2460            return extval;
   2461        }
   2462    } else if (shift < 48) {
   2463        int64_t extval = sextract64(src << shift, 0, 48);
   2464        if (!sat || src == (extval >> shift)) {
   2465            return extval;
   2466        }
   2467    } else if (!sat || src == 0) {
   2468        return 0;
   2469    }
   2470
   2471    *sat = 1;
   2472    return src >= 0 ? MAKE_64BIT_MASK(0, 47) : MAKE_64BIT_MASK(47, 17);
   2473}
   2474
   2475/* Operate on 64-bit values, but saturate at 48 bits */
   2476static inline uint64_t do_uqrshl48_d(uint64_t src, int64_t shift,
   2477                                     bool round, uint32_t *sat)
   2478{
   2479    uint64_t val, extval;
   2480
   2481    if (shift <= -(48 + round)) {
   2482        return 0;
   2483    } else if (shift < 0) {
   2484        if (round) {
   2485            val = src >> (-shift - 1);
   2486            val = (val >> 1) + (val & 1);
   2487        } else {
   2488            val = src >> -shift;
   2489        }
   2490        extval = extract64(val, 0, 48);
   2491        if (!sat || val == extval) {
   2492            return extval;
   2493        }
   2494    } else if (shift < 48) {
   2495        uint64_t extval = extract64(src << shift, 0, 48);
   2496        if (!sat || src == (extval >> shift)) {
   2497            return extval;
   2498        }
   2499    } else if (!sat || src == 0) {
   2500        return 0;
   2501    }
   2502
   2503    *sat = 1;
   2504    return MAKE_64BIT_MASK(0, 48);
   2505}
   2506
   2507uint64_t HELPER(mve_sqrshrl48)(CPUARMState *env, uint64_t n, uint32_t shift)
   2508{
   2509    return do_sqrshl48_d(n, -(int8_t)shift, true, &env->QF);
   2510}
   2511
   2512uint64_t HELPER(mve_uqrshll48)(CPUARMState *env, uint64_t n, uint32_t shift)
   2513{
   2514    return do_uqrshl48_d(n, (int8_t)shift, true, &env->QF);
   2515}
   2516
   2517uint32_t HELPER(mve_uqshl)(CPUARMState *env, uint32_t n, uint32_t shift)
   2518{
   2519    return do_uqrshl_bhs(n, (int8_t)shift, 32, false, &env->QF);
   2520}
   2521
   2522uint32_t HELPER(mve_sqshl)(CPUARMState *env, uint32_t n, uint32_t shift)
   2523{
   2524    return do_sqrshl_bhs(n, (int8_t)shift, 32, false, &env->QF);
   2525}
   2526
   2527uint32_t HELPER(mve_uqrshl)(CPUARMState *env, uint32_t n, uint32_t shift)
   2528{
   2529    return do_uqrshl_bhs(n, (int8_t)shift, 32, true, &env->QF);
   2530}
   2531
   2532uint32_t HELPER(mve_sqrshr)(CPUARMState *env, uint32_t n, uint32_t shift)
   2533{
   2534    return do_sqrshl_bhs(n, -(int8_t)shift, 32, true, &env->QF);
   2535}
   2536
   2537#define DO_VIDUP(OP, ESIZE, TYPE, FN)                           \
   2538    uint32_t HELPER(mve_##OP)(CPUARMState *env, void *vd,       \
   2539                           uint32_t offset, uint32_t imm)       \
   2540    {                                                           \
   2541        TYPE *d = vd;                                           \
   2542        uint16_t mask = mve_element_mask(env);                  \
   2543        unsigned e;                                             \
   2544        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   2545            mergemask(&d[H##ESIZE(e)], offset, mask);           \
   2546            offset = FN(offset, imm);                           \
   2547        }                                                       \
   2548        mve_advance_vpt(env);                                   \
   2549        return offset;                                          \
   2550    }
   2551
   2552#define DO_VIWDUP(OP, ESIZE, TYPE, FN)                          \
   2553    uint32_t HELPER(mve_##OP)(CPUARMState *env, void *vd,       \
   2554                              uint32_t offset, uint32_t wrap,   \
   2555                              uint32_t imm)                     \
   2556    {                                                           \
   2557        TYPE *d = vd;                                           \
   2558        uint16_t mask = mve_element_mask(env);                  \
   2559        unsigned e;                                             \
   2560        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   2561            mergemask(&d[H##ESIZE(e)], offset, mask);           \
   2562            offset = FN(offset, wrap, imm);                     \
   2563        }                                                       \
   2564        mve_advance_vpt(env);                                   \
   2565        return offset;                                          \
   2566    }
   2567
   2568#define DO_VIDUP_ALL(OP, FN)                    \
   2569    DO_VIDUP(OP##b, 1, int8_t, FN)              \
   2570    DO_VIDUP(OP##h, 2, int16_t, FN)             \
   2571    DO_VIDUP(OP##w, 4, int32_t, FN)
   2572
   2573#define DO_VIWDUP_ALL(OP, FN)                   \
   2574    DO_VIWDUP(OP##b, 1, int8_t, FN)             \
   2575    DO_VIWDUP(OP##h, 2, int16_t, FN)            \
   2576    DO_VIWDUP(OP##w, 4, int32_t, FN)
   2577
   2578static uint32_t do_add_wrap(uint32_t offset, uint32_t wrap, uint32_t imm)
   2579{
   2580    offset += imm;
   2581    if (offset == wrap) {
   2582        offset = 0;
   2583    }
   2584    return offset;
   2585}
   2586
   2587static uint32_t do_sub_wrap(uint32_t offset, uint32_t wrap, uint32_t imm)
   2588{
   2589    if (offset == 0) {
   2590        offset = wrap;
   2591    }
   2592    offset -= imm;
   2593    return offset;
   2594}
   2595
   2596DO_VIDUP_ALL(vidup, DO_ADD)
   2597DO_VIWDUP_ALL(viwdup, do_add_wrap)
   2598DO_VIWDUP_ALL(vdwdup, do_sub_wrap)
   2599
   2600/*
   2601 * Vector comparison.
   2602 * P0 bits for non-executed beats (where eci_mask is 0) are unchanged.
   2603 * P0 bits for predicated lanes in executed beats (where mask is 0) are 0.
   2604 * P0 bits otherwise are updated with the results of the comparisons.
   2605 * We must also keep unchanged the MASK fields at the top of v7m.vpr.
   2606 */
   2607#define DO_VCMP(OP, ESIZE, TYPE, FN)                                    \
   2608    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vn, void *vm)   \
   2609    {                                                                   \
   2610        TYPE *n = vn, *m = vm;                                          \
   2611        uint16_t mask = mve_element_mask(env);                          \
   2612        uint16_t eci_mask = mve_eci_mask(env);                          \
   2613        uint16_t beatpred = 0;                                          \
   2614        uint16_t emask = MAKE_64BIT_MASK(0, ESIZE);                     \
   2615        unsigned e;                                                     \
   2616        for (e = 0; e < 16 / ESIZE; e++) {                              \
   2617            bool r = FN(n[H##ESIZE(e)], m[H##ESIZE(e)]);                \
   2618            /* Comparison sets 0/1 bits for each byte in the element */ \
   2619            beatpred |= r * emask;                                      \
   2620            emask <<= ESIZE;                                            \
   2621        }                                                               \
   2622        beatpred &= mask;                                               \
   2623        env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) |           \
   2624            (beatpred & eci_mask);                                      \
   2625        mve_advance_vpt(env);                                           \
   2626    }
   2627
   2628#define DO_VCMP_SCALAR(OP, ESIZE, TYPE, FN)                             \
   2629    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vn,             \
   2630                                uint32_t rm)                            \
   2631    {                                                                   \
   2632        TYPE *n = vn;                                                   \
   2633        uint16_t mask = mve_element_mask(env);                          \
   2634        uint16_t eci_mask = mve_eci_mask(env);                          \
   2635        uint16_t beatpred = 0;                                          \
   2636        uint16_t emask = MAKE_64BIT_MASK(0, ESIZE);                     \
   2637        unsigned e;                                                     \
   2638        for (e = 0; e < 16 / ESIZE; e++) {                              \
   2639            bool r = FN(n[H##ESIZE(e)], (TYPE)rm);                      \
   2640            /* Comparison sets 0/1 bits for each byte in the element */ \
   2641            beatpred |= r * emask;                                      \
   2642            emask <<= ESIZE;                                            \
   2643        }                                                               \
   2644        beatpred &= mask;                                               \
   2645        env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) |           \
   2646            (beatpred & eci_mask);                                      \
   2647        mve_advance_vpt(env);                                           \
   2648    }
   2649
   2650#define DO_VCMP_S(OP, FN)                               \
   2651    DO_VCMP(OP##b, 1, int8_t, FN)                       \
   2652    DO_VCMP(OP##h, 2, int16_t, FN)                      \
   2653    DO_VCMP(OP##w, 4, int32_t, FN)                      \
   2654    DO_VCMP_SCALAR(OP##_scalarb, 1, int8_t, FN)         \
   2655    DO_VCMP_SCALAR(OP##_scalarh, 2, int16_t, FN)        \
   2656    DO_VCMP_SCALAR(OP##_scalarw, 4, int32_t, FN)
   2657
   2658#define DO_VCMP_U(OP, FN)                               \
   2659    DO_VCMP(OP##b, 1, uint8_t, FN)                      \
   2660    DO_VCMP(OP##h, 2, uint16_t, FN)                     \
   2661    DO_VCMP(OP##w, 4, uint32_t, FN)                     \
   2662    DO_VCMP_SCALAR(OP##_scalarb, 1, uint8_t, FN)        \
   2663    DO_VCMP_SCALAR(OP##_scalarh, 2, uint16_t, FN)       \
   2664    DO_VCMP_SCALAR(OP##_scalarw, 4, uint32_t, FN)
   2665
   2666#define DO_EQ(N, M) ((N) == (M))
   2667#define DO_NE(N, M) ((N) != (M))
   2668#define DO_EQ(N, M) ((N) == (M))
   2669#define DO_EQ(N, M) ((N) == (M))
   2670#define DO_GE(N, M) ((N) >= (M))
   2671#define DO_LT(N, M) ((N) < (M))
   2672#define DO_GT(N, M) ((N) > (M))
   2673#define DO_LE(N, M) ((N) <= (M))
   2674
   2675DO_VCMP_U(vcmpeq, DO_EQ)
   2676DO_VCMP_U(vcmpne, DO_NE)
   2677DO_VCMP_U(vcmpcs, DO_GE)
   2678DO_VCMP_U(vcmphi, DO_GT)
   2679DO_VCMP_S(vcmpge, DO_GE)
   2680DO_VCMP_S(vcmplt, DO_LT)
   2681DO_VCMP_S(vcmpgt, DO_GT)
   2682DO_VCMP_S(vcmple, DO_LE)
   2683
   2684void HELPER(mve_vpsel)(CPUARMState *env, void *vd, void *vn, void *vm)
   2685{
   2686    /*
   2687     * Qd[n] = VPR.P0[n] ? Qn[n] : Qm[n]
   2688     * but note that whether bytes are written to Qd is still subject
   2689     * to (all forms of) predication in the usual way.
   2690     */
   2691    uint64_t *d = vd, *n = vn, *m = vm;
   2692    uint16_t mask = mve_element_mask(env);
   2693    uint16_t p0 = FIELD_EX32(env->v7m.vpr, V7M_VPR, P0);
   2694    unsigned e;
   2695    for (e = 0; e < 16 / 8; e++, mask >>= 8, p0 >>= 8) {
   2696        uint64_t r = m[H8(e)];
   2697        mergemask(&r, n[H8(e)], p0);
   2698        mergemask(&d[H8(e)], r, mask);
   2699    }
   2700    mve_advance_vpt(env);
   2701}
   2702
   2703void HELPER(mve_vpnot)(CPUARMState *env)
   2704{
   2705    /*
   2706     * P0 bits for unexecuted beats (where eci_mask is 0) are unchanged.
   2707     * P0 bits for predicated lanes in executed bits (where mask is 0) are 0.
   2708     * P0 bits otherwise are inverted.
   2709     * (This is the same logic as VCMP.)
   2710     * This insn is itself subject to predication and to beat-wise execution,
   2711     * and after it executes VPT state advances in the usual way.
   2712     */
   2713    uint16_t mask = mve_element_mask(env);
   2714    uint16_t eci_mask = mve_eci_mask(env);
   2715    uint16_t beatpred = ~env->v7m.vpr & mask;
   2716    env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) | (beatpred & eci_mask);
   2717    mve_advance_vpt(env);
   2718}
   2719
   2720/*
   2721 * VCTP: P0 unexecuted bits unchanged, predicated bits zeroed,
   2722 * otherwise set according to value of Rn. The calculation of
   2723 * newmask here works in the same way as the calculation of the
   2724 * ltpmask in mve_element_mask(), but we have pre-calculated
   2725 * the masklen in the generated code.
   2726 */
   2727void HELPER(mve_vctp)(CPUARMState *env, uint32_t masklen)
   2728{
   2729    uint16_t mask = mve_element_mask(env);
   2730    uint16_t eci_mask = mve_eci_mask(env);
   2731    uint16_t newmask;
   2732
   2733    assert(masklen <= 16);
   2734    newmask = masklen ? MAKE_64BIT_MASK(0, masklen) : 0;
   2735    newmask &= mask;
   2736    env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) | (newmask & eci_mask);
   2737    mve_advance_vpt(env);
   2738}
   2739
   2740#define DO_1OP_SAT(OP, ESIZE, TYPE, FN)                                 \
   2741    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm)         \
   2742    {                                                                   \
   2743        TYPE *d = vd, *m = vm;                                          \
   2744        uint16_t mask = mve_element_mask(env);                          \
   2745        unsigned e;                                                     \
   2746        bool qc = false;                                                \
   2747        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   2748            bool sat = false;                                           \
   2749            mergemask(&d[H##ESIZE(e)], FN(m[H##ESIZE(e)], &sat), mask); \
   2750            qc |= sat & mask & 1;                                       \
   2751        }                                                               \
   2752        if (qc) {                                                       \
   2753            env->vfp.qc[0] = qc;                                        \
   2754        }                                                               \
   2755        mve_advance_vpt(env);                                           \
   2756    }
   2757
   2758#define DO_VQABS_B(N, SATP) \
   2759    do_sat_bhs(DO_ABS((int64_t)N), INT8_MIN, INT8_MAX, SATP)
   2760#define DO_VQABS_H(N, SATP) \
   2761    do_sat_bhs(DO_ABS((int64_t)N), INT16_MIN, INT16_MAX, SATP)
   2762#define DO_VQABS_W(N, SATP) \
   2763    do_sat_bhs(DO_ABS((int64_t)N), INT32_MIN, INT32_MAX, SATP)
   2764
   2765#define DO_VQNEG_B(N, SATP) do_sat_bhs(-(int64_t)N, INT8_MIN, INT8_MAX, SATP)
   2766#define DO_VQNEG_H(N, SATP) do_sat_bhs(-(int64_t)N, INT16_MIN, INT16_MAX, SATP)
   2767#define DO_VQNEG_W(N, SATP) do_sat_bhs(-(int64_t)N, INT32_MIN, INT32_MAX, SATP)
   2768
   2769DO_1OP_SAT(vqabsb, 1, int8_t, DO_VQABS_B)
   2770DO_1OP_SAT(vqabsh, 2, int16_t, DO_VQABS_H)
   2771DO_1OP_SAT(vqabsw, 4, int32_t, DO_VQABS_W)
   2772
   2773DO_1OP_SAT(vqnegb, 1, int8_t, DO_VQNEG_B)
   2774DO_1OP_SAT(vqnegh, 2, int16_t, DO_VQNEG_H)
   2775DO_1OP_SAT(vqnegw, 4, int32_t, DO_VQNEG_W)
   2776
   2777/*
   2778 * VMAXA, VMINA: vd is unsigned; vm is signed, and we take its
   2779 * absolute value; we then do an unsigned comparison.
   2780 */
   2781#define DO_VMAXMINA(OP, ESIZE, STYPE, UTYPE, FN)                        \
   2782    void HELPER(mve_##OP)(CPUARMState *env, void *vd, void *vm)         \
   2783    {                                                                   \
   2784        UTYPE *d = vd;                                                  \
   2785        STYPE *m = vm;                                                  \
   2786        uint16_t mask = mve_element_mask(env);                          \
   2787        unsigned e;                                                     \
   2788        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   2789            UTYPE r = DO_ABS(m[H##ESIZE(e)]);                           \
   2790            r = FN(d[H##ESIZE(e)], r);                                  \
   2791            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   2792        }                                                               \
   2793        mve_advance_vpt(env);                                           \
   2794    }
   2795
   2796DO_VMAXMINA(vmaxab, 1, int8_t, uint8_t, DO_MAX)
   2797DO_VMAXMINA(vmaxah, 2, int16_t, uint16_t, DO_MAX)
   2798DO_VMAXMINA(vmaxaw, 4, int32_t, uint32_t, DO_MAX)
   2799DO_VMAXMINA(vminab, 1, int8_t, uint8_t, DO_MIN)
   2800DO_VMAXMINA(vminah, 2, int16_t, uint16_t, DO_MIN)
   2801DO_VMAXMINA(vminaw, 4, int32_t, uint32_t, DO_MIN)
   2802
   2803/*
   2804 * 2-operand floating point. Note that if an element is partially
   2805 * predicated we must do the FP operation to update the non-predicated
   2806 * bytes, but we must be careful to avoid updating the FP exception
   2807 * state unless byte 0 of the element was unpredicated.
   2808 */
   2809#define DO_2OP_FP(OP, ESIZE, TYPE, FN)                                  \
   2810    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   2811                                void *vd, void *vn, void *vm)           \
   2812    {                                                                   \
   2813        TYPE *d = vd, *n = vn, *m = vm;                                 \
   2814        TYPE r;                                                         \
   2815        uint16_t mask = mve_element_mask(env);                          \
   2816        unsigned e;                                                     \
   2817        float_status *fpst;                                             \
   2818        float_status scratch_fpst;                                      \
   2819        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   2820            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   2821                continue;                                               \
   2822            }                                                           \
   2823            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   2824                &env->vfp.standard_fp_status;                           \
   2825            if (!(mask & 1)) {                                          \
   2826                /* We need the result but without updating flags */     \
   2827                scratch_fpst = *fpst;                                   \
   2828                fpst = &scratch_fpst;                                   \
   2829            }                                                           \
   2830            r = FN(n[H##ESIZE(e)], m[H##ESIZE(e)], fpst);               \
   2831            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   2832        }                                                               \
   2833        mve_advance_vpt(env);                                           \
   2834    }
   2835
   2836#define DO_2OP_FP_ALL(OP, FN)                  \
   2837    DO_2OP_FP(OP##h, 2, float16, float16_##FN) \
   2838    DO_2OP_FP(OP##s, 4, float32, float32_##FN)
   2839
   2840DO_2OP_FP_ALL(vfadd, add)
   2841DO_2OP_FP_ALL(vfsub, sub)
   2842DO_2OP_FP_ALL(vfmul, mul)
   2843
   2844static inline float16 float16_abd(float16 a, float16 b, float_status *s)
   2845{
   2846    return float16_abs(float16_sub(a, b, s));
   2847}
   2848
   2849static inline float32 float32_abd(float32 a, float32 b, float_status *s)
   2850{
   2851    return float32_abs(float32_sub(a, b, s));
   2852}
   2853
   2854DO_2OP_FP_ALL(vfabd, abd)
   2855DO_2OP_FP_ALL(vmaxnm, maxnum)
   2856DO_2OP_FP_ALL(vminnm, minnum)
   2857
   2858static inline float16 float16_maxnuma(float16 a, float16 b, float_status *s)
   2859{
   2860    return float16_maxnum(float16_abs(a), float16_abs(b), s);
   2861}
   2862
   2863static inline float32 float32_maxnuma(float32 a, float32 b, float_status *s)
   2864{
   2865    return float32_maxnum(float32_abs(a), float32_abs(b), s);
   2866}
   2867
   2868static inline float16 float16_minnuma(float16 a, float16 b, float_status *s)
   2869{
   2870    return float16_minnum(float16_abs(a), float16_abs(b), s);
   2871}
   2872
   2873static inline float32 float32_minnuma(float32 a, float32 b, float_status *s)
   2874{
   2875    return float32_minnum(float32_abs(a), float32_abs(b), s);
   2876}
   2877
   2878DO_2OP_FP_ALL(vmaxnma, maxnuma)
   2879DO_2OP_FP_ALL(vminnma, minnuma)
   2880
   2881#define DO_VCADD_FP(OP, ESIZE, TYPE, FN0, FN1)                          \
   2882    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   2883                                void *vd, void *vn, void *vm)           \
   2884    {                                                                   \
   2885        TYPE *d = vd, *n = vn, *m = vm;                                 \
   2886        TYPE r[16 / ESIZE];                                             \
   2887        uint16_t tm, mask = mve_element_mask(env);                      \
   2888        unsigned e;                                                     \
   2889        float_status *fpst;                                             \
   2890        float_status scratch_fpst;                                      \
   2891        /* Calculate all results first to avoid overwriting inputs */   \
   2892        for (e = 0, tm = mask; e < 16 / ESIZE; e++, tm >>= ESIZE) {     \
   2893            if ((tm & MAKE_64BIT_MASK(0, ESIZE)) == 0) {                \
   2894                r[e] = 0;                                               \
   2895                continue;                                               \
   2896            }                                                           \
   2897            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   2898                &env->vfp.standard_fp_status;                           \
   2899            if (!(tm & 1)) {                                            \
   2900                /* We need the result but without updating flags */     \
   2901                scratch_fpst = *fpst;                                   \
   2902                fpst = &scratch_fpst;                                   \
   2903            }                                                           \
   2904            if (!(e & 1)) {                                             \
   2905                r[e] = FN0(n[H##ESIZE(e)], m[H##ESIZE(e + 1)], fpst);   \
   2906            } else {                                                    \
   2907                r[e] = FN1(n[H##ESIZE(e)], m[H##ESIZE(e - 1)], fpst);   \
   2908            }                                                           \
   2909        }                                                               \
   2910        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   2911            mergemask(&d[H##ESIZE(e)], r[e], mask);                     \
   2912        }                                                               \
   2913        mve_advance_vpt(env);                                           \
   2914    }
   2915
   2916DO_VCADD_FP(vfcadd90h, 2, float16, float16_sub, float16_add)
   2917DO_VCADD_FP(vfcadd90s, 4, float32, float32_sub, float32_add)
   2918DO_VCADD_FP(vfcadd270h, 2, float16, float16_add, float16_sub)
   2919DO_VCADD_FP(vfcadd270s, 4, float32, float32_add, float32_sub)
   2920
   2921#define DO_VFMA(OP, ESIZE, TYPE, CHS)                                   \
   2922    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   2923                                void *vd, void *vn, void *vm)           \
   2924    {                                                                   \
   2925        TYPE *d = vd, *n = vn, *m = vm;                                 \
   2926        TYPE r;                                                         \
   2927        uint16_t mask = mve_element_mask(env);                          \
   2928        unsigned e;                                                     \
   2929        float_status *fpst;                                             \
   2930        float_status scratch_fpst;                                      \
   2931        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   2932            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   2933                continue;                                               \
   2934            }                                                           \
   2935            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   2936                &env->vfp.standard_fp_status;                           \
   2937            if (!(mask & 1)) {                                          \
   2938                /* We need the result but without updating flags */     \
   2939                scratch_fpst = *fpst;                                   \
   2940                fpst = &scratch_fpst;                                   \
   2941            }                                                           \
   2942            r = n[H##ESIZE(e)];                                         \
   2943            if (CHS) {                                                  \
   2944                r = TYPE##_chs(r);                                      \
   2945            }                                                           \
   2946            r = TYPE##_muladd(r, m[H##ESIZE(e)], d[H##ESIZE(e)],        \
   2947                              0, fpst);                                 \
   2948            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   2949        }                                                               \
   2950        mve_advance_vpt(env);                                           \
   2951    }
   2952
   2953DO_VFMA(vfmah, 2, float16, false)
   2954DO_VFMA(vfmas, 4, float32, false)
   2955DO_VFMA(vfmsh, 2, float16, true)
   2956DO_VFMA(vfmss, 4, float32, true)
   2957
   2958#define DO_VCMLA(OP, ESIZE, TYPE, ROT, FN)                              \
   2959    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   2960                                void *vd, void *vn, void *vm)           \
   2961    {                                                                   \
   2962        TYPE *d = vd, *n = vn, *m = vm;                                 \
   2963        TYPE r0, r1, e1, e2, e3, e4;                                    \
   2964        uint16_t mask = mve_element_mask(env);                          \
   2965        unsigned e;                                                     \
   2966        float_status *fpst0, *fpst1;                                    \
   2967        float_status scratch_fpst;                                      \
   2968        /* We loop through pairs of elements at a time */               \
   2969        for (e = 0; e < 16 / ESIZE; e += 2, mask >>= ESIZE * 2) {       \
   2970            if ((mask & MAKE_64BIT_MASK(0, ESIZE * 2)) == 0) {          \
   2971                continue;                                               \
   2972            }                                                           \
   2973            fpst0 = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :   \
   2974                &env->vfp.standard_fp_status;                           \
   2975            fpst1 = fpst0;                                              \
   2976            if (!(mask & 1)) {                                          \
   2977                scratch_fpst = *fpst0;                                  \
   2978                fpst0 = &scratch_fpst;                                  \
   2979            }                                                           \
   2980            if (!(mask & (1 << ESIZE))) {                               \
   2981                scratch_fpst = *fpst1;                                  \
   2982                fpst1 = &scratch_fpst;                                  \
   2983            }                                                           \
   2984            switch (ROT) {                                              \
   2985            case 0:                                                     \
   2986                e1 = m[H##ESIZE(e)];                                    \
   2987                e2 = n[H##ESIZE(e)];                                    \
   2988                e3 = m[H##ESIZE(e + 1)];                                \
   2989                e4 = n[H##ESIZE(e)];                                    \
   2990                break;                                                  \
   2991            case 1:                                                     \
   2992                e1 = TYPE##_chs(m[H##ESIZE(e + 1)]);                    \
   2993                e2 = n[H##ESIZE(e + 1)];                                \
   2994                e3 = m[H##ESIZE(e)];                                    \
   2995                e4 = n[H##ESIZE(e + 1)];                                \
   2996                break;                                                  \
   2997            case 2:                                                     \
   2998                e1 = TYPE##_chs(m[H##ESIZE(e)]);                        \
   2999                e2 = n[H##ESIZE(e)];                                    \
   3000                e3 = TYPE##_chs(m[H##ESIZE(e + 1)]);                    \
   3001                e4 = n[H##ESIZE(e)];                                    \
   3002                break;                                                  \
   3003            case 3:                                                     \
   3004                e1 = m[H##ESIZE(e + 1)];                                \
   3005                e2 = n[H##ESIZE(e + 1)];                                \
   3006                e3 = TYPE##_chs(m[H##ESIZE(e)]);                        \
   3007                e4 = n[H##ESIZE(e + 1)];                                \
   3008                break;                                                  \
   3009            default:                                                    \
   3010                g_assert_not_reached();                                 \
   3011            }                                                           \
   3012            r0 = FN(e2, e1, d[H##ESIZE(e)], fpst0);                     \
   3013            r1 = FN(e4, e3, d[H##ESIZE(e + 1)], fpst1);                 \
   3014            mergemask(&d[H##ESIZE(e)], r0, mask);                       \
   3015            mergemask(&d[H##ESIZE(e + 1)], r1, mask >> ESIZE);          \
   3016        }                                                               \
   3017        mve_advance_vpt(env);                                           \
   3018    }
   3019
   3020#define DO_VCMULH(N, M, D, S) float16_mul(N, M, S)
   3021#define DO_VCMULS(N, M, D, S) float32_mul(N, M, S)
   3022
   3023#define DO_VCMLAH(N, M, D, S) float16_muladd(N, M, D, 0, S)
   3024#define DO_VCMLAS(N, M, D, S) float32_muladd(N, M, D, 0, S)
   3025
   3026DO_VCMLA(vcmul0h, 2, float16, 0, DO_VCMULH)
   3027DO_VCMLA(vcmul0s, 4, float32, 0, DO_VCMULS)
   3028DO_VCMLA(vcmul90h, 2, float16, 1, DO_VCMULH)
   3029DO_VCMLA(vcmul90s, 4, float32, 1, DO_VCMULS)
   3030DO_VCMLA(vcmul180h, 2, float16, 2, DO_VCMULH)
   3031DO_VCMLA(vcmul180s, 4, float32, 2, DO_VCMULS)
   3032DO_VCMLA(vcmul270h, 2, float16, 3, DO_VCMULH)
   3033DO_VCMLA(vcmul270s, 4, float32, 3, DO_VCMULS)
   3034
   3035DO_VCMLA(vcmla0h, 2, float16, 0, DO_VCMLAH)
   3036DO_VCMLA(vcmla0s, 4, float32, 0, DO_VCMLAS)
   3037DO_VCMLA(vcmla90h, 2, float16, 1, DO_VCMLAH)
   3038DO_VCMLA(vcmla90s, 4, float32, 1, DO_VCMLAS)
   3039DO_VCMLA(vcmla180h, 2, float16, 2, DO_VCMLAH)
   3040DO_VCMLA(vcmla180s, 4, float32, 2, DO_VCMLAS)
   3041DO_VCMLA(vcmla270h, 2, float16, 3, DO_VCMLAH)
   3042DO_VCMLA(vcmla270s, 4, float32, 3, DO_VCMLAS)
   3043
   3044#define DO_2OP_FP_SCALAR(OP, ESIZE, TYPE, FN)                           \
   3045    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   3046                                void *vd, void *vn, uint32_t rm)        \
   3047    {                                                                   \
   3048        TYPE *d = vd, *n = vn;                                          \
   3049        TYPE r, m = rm;                                                 \
   3050        uint16_t mask = mve_element_mask(env);                          \
   3051        unsigned e;                                                     \
   3052        float_status *fpst;                                             \
   3053        float_status scratch_fpst;                                      \
   3054        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   3055            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   3056                continue;                                               \
   3057            }                                                           \
   3058            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3059                &env->vfp.standard_fp_status;                           \
   3060            if (!(mask & 1)) {                                          \
   3061                /* We need the result but without updating flags */     \
   3062                scratch_fpst = *fpst;                                   \
   3063                fpst = &scratch_fpst;                                   \
   3064            }                                                           \
   3065            r = FN(n[H##ESIZE(e)], m, fpst);                            \
   3066            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   3067        }                                                               \
   3068        mve_advance_vpt(env);                                           \
   3069    }
   3070
   3071#define DO_2OP_FP_SCALAR_ALL(OP, FN)                    \
   3072    DO_2OP_FP_SCALAR(OP##h, 2, float16, float16_##FN)   \
   3073    DO_2OP_FP_SCALAR(OP##s, 4, float32, float32_##FN)
   3074
   3075DO_2OP_FP_SCALAR_ALL(vfadd_scalar, add)
   3076DO_2OP_FP_SCALAR_ALL(vfsub_scalar, sub)
   3077DO_2OP_FP_SCALAR_ALL(vfmul_scalar, mul)
   3078
   3079#define DO_2OP_FP_ACC_SCALAR(OP, ESIZE, TYPE, FN)                       \
   3080    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   3081                                void *vd, void *vn, uint32_t rm)        \
   3082    {                                                                   \
   3083        TYPE *d = vd, *n = vn;                                          \
   3084        TYPE r, m = rm;                                                 \
   3085        uint16_t mask = mve_element_mask(env);                          \
   3086        unsigned e;                                                     \
   3087        float_status *fpst;                                             \
   3088        float_status scratch_fpst;                                      \
   3089        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   3090            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   3091                continue;                                               \
   3092            }                                                           \
   3093            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3094                &env->vfp.standard_fp_status;                           \
   3095            if (!(mask & 1)) {                                          \
   3096                /* We need the result but without updating flags */     \
   3097                scratch_fpst = *fpst;                                   \
   3098                fpst = &scratch_fpst;                                   \
   3099            }                                                           \
   3100            r = FN(n[H##ESIZE(e)], m, d[H##ESIZE(e)], 0, fpst);         \
   3101            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   3102        }                                                               \
   3103        mve_advance_vpt(env);                                           \
   3104    }
   3105
   3106/* VFMAS is vector * vector + scalar, so swap op2 and op3 */
   3107#define DO_VFMAS_SCALARH(N, M, D, F, S) float16_muladd(N, D, M, F, S)
   3108#define DO_VFMAS_SCALARS(N, M, D, F, S) float32_muladd(N, D, M, F, S)
   3109
   3110/* VFMA is vector * scalar + vector */
   3111DO_2OP_FP_ACC_SCALAR(vfma_scalarh, 2, float16, float16_muladd)
   3112DO_2OP_FP_ACC_SCALAR(vfma_scalars, 4, float32, float32_muladd)
   3113DO_2OP_FP_ACC_SCALAR(vfmas_scalarh, 2, float16, DO_VFMAS_SCALARH)
   3114DO_2OP_FP_ACC_SCALAR(vfmas_scalars, 4, float32, DO_VFMAS_SCALARS)
   3115
   3116/* Floating point max/min across vector. */
   3117#define DO_FP_VMAXMINV(OP, ESIZE, TYPE, ABS, FN)                \
   3118    uint32_t HELPER(glue(mve_, OP))(CPUARMState *env, void *vm, \
   3119                                    uint32_t ra_in)             \
   3120    {                                                           \
   3121        uint16_t mask = mve_element_mask(env);                  \
   3122        unsigned e;                                             \
   3123        TYPE *m = vm;                                           \
   3124        TYPE ra = (TYPE)ra_in;                                  \
   3125        float_status *fpst = (ESIZE == 2) ?                     \
   3126            &env->vfp.standard_fp_status_f16 :                  \
   3127            &env->vfp.standard_fp_status;                       \
   3128        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {      \
   3129            if (mask & 1) {                                     \
   3130                TYPE v = m[H##ESIZE(e)];                        \
   3131                if (TYPE##_is_signaling_nan(ra, fpst)) {        \
   3132                    ra = TYPE##_silence_nan(ra, fpst);          \
   3133                    float_raise(float_flag_invalid, fpst);      \
   3134                }                                               \
   3135                if (TYPE##_is_signaling_nan(v, fpst)) {         \
   3136                    v = TYPE##_silence_nan(v, fpst);            \
   3137                    float_raise(float_flag_invalid, fpst);      \
   3138                }                                               \
   3139                if (ABS) {                                      \
   3140                    v = TYPE##_abs(v);                          \
   3141                }                                               \
   3142                ra = FN(ra, v, fpst);                           \
   3143            }                                                   \
   3144        }                                                       \
   3145        mve_advance_vpt(env);                                   \
   3146        return ra;                                              \
   3147    }                                                           \
   3148
   3149#define NOP(X) (X)
   3150
   3151DO_FP_VMAXMINV(vmaxnmvh, 2, float16, false, float16_maxnum)
   3152DO_FP_VMAXMINV(vmaxnmvs, 4, float32, false, float32_maxnum)
   3153DO_FP_VMAXMINV(vminnmvh, 2, float16, false, float16_minnum)
   3154DO_FP_VMAXMINV(vminnmvs, 4, float32, false, float32_minnum)
   3155DO_FP_VMAXMINV(vmaxnmavh, 2, float16, true, float16_maxnum)
   3156DO_FP_VMAXMINV(vmaxnmavs, 4, float32, true, float32_maxnum)
   3157DO_FP_VMAXMINV(vminnmavh, 2, float16, true, float16_minnum)
   3158DO_FP_VMAXMINV(vminnmavs, 4, float32, true, float32_minnum)
   3159
   3160/* FP compares; note that all comparisons signal InvalidOp for QNaNs */
   3161#define DO_VCMP_FP(OP, ESIZE, TYPE, FN)                                 \
   3162    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vn, void *vm)   \
   3163    {                                                                   \
   3164        TYPE *n = vn, *m = vm;                                          \
   3165        uint16_t mask = mve_element_mask(env);                          \
   3166        uint16_t eci_mask = mve_eci_mask(env);                          \
   3167        uint16_t beatpred = 0;                                          \
   3168        uint16_t emask = MAKE_64BIT_MASK(0, ESIZE);                     \
   3169        unsigned e;                                                     \
   3170        float_status *fpst;                                             \
   3171        float_status scratch_fpst;                                      \
   3172        bool r;                                                         \
   3173        for (e = 0; e < 16 / ESIZE; e++, emask <<= ESIZE) {             \
   3174            if ((mask & emask) == 0) {                                  \
   3175                continue;                                               \
   3176            }                                                           \
   3177            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3178                &env->vfp.standard_fp_status;                           \
   3179            if (!(mask & (1 << (e * ESIZE)))) {                         \
   3180                /* We need the result but without updating flags */     \
   3181                scratch_fpst = *fpst;                                   \
   3182                fpst = &scratch_fpst;                                   \
   3183            }                                                           \
   3184            r = FN(n[H##ESIZE(e)], m[H##ESIZE(e)], fpst);               \
   3185            /* Comparison sets 0/1 bits for each byte in the element */ \
   3186            beatpred |= r * emask;                                      \
   3187        }                                                               \
   3188        beatpred &= mask;                                               \
   3189        env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) |           \
   3190            (beatpred & eci_mask);                                      \
   3191        mve_advance_vpt(env);                                           \
   3192    }
   3193
   3194#define DO_VCMP_FP_SCALAR(OP, ESIZE, TYPE, FN)                          \
   3195    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vn,             \
   3196                                uint32_t rm)                            \
   3197    {                                                                   \
   3198        TYPE *n = vn;                                                   \
   3199        uint16_t mask = mve_element_mask(env);                          \
   3200        uint16_t eci_mask = mve_eci_mask(env);                          \
   3201        uint16_t beatpred = 0;                                          \
   3202        uint16_t emask = MAKE_64BIT_MASK(0, ESIZE);                     \
   3203        unsigned e;                                                     \
   3204        float_status *fpst;                                             \
   3205        float_status scratch_fpst;                                      \
   3206        bool r;                                                         \
   3207        for (e = 0; e < 16 / ESIZE; e++, emask <<= ESIZE) {             \
   3208            if ((mask & emask) == 0) {                                  \
   3209                continue;                                               \
   3210            }                                                           \
   3211            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3212                &env->vfp.standard_fp_status;                           \
   3213            if (!(mask & (1 << (e * ESIZE)))) {                         \
   3214                /* We need the result but without updating flags */     \
   3215                scratch_fpst = *fpst;                                   \
   3216                fpst = &scratch_fpst;                                   \
   3217            }                                                           \
   3218            r = FN(n[H##ESIZE(e)], (TYPE)rm, fpst);                     \
   3219            /* Comparison sets 0/1 bits for each byte in the element */ \
   3220            beatpred |= r * emask;                                      \
   3221        }                                                               \
   3222        beatpred &= mask;                                               \
   3223        env->v7m.vpr = (env->v7m.vpr & ~(uint32_t)eci_mask) |           \
   3224            (beatpred & eci_mask);                                      \
   3225        mve_advance_vpt(env);                                           \
   3226    }
   3227
   3228#define DO_VCMP_FP_BOTH(VOP, SOP, ESIZE, TYPE, FN)      \
   3229    DO_VCMP_FP(VOP, ESIZE, TYPE, FN)                    \
   3230    DO_VCMP_FP_SCALAR(SOP, ESIZE, TYPE, FN)
   3231
   3232/*
   3233 * Some care is needed here to get the correct result for the unordered case.
   3234 * Architecturally EQ, GE and GT are defined to be false for unordered, but
   3235 * the NE, LT and LE comparisons are defined as simple logical inverses of
   3236 * EQ, GE and GT and so they must return true for unordered. The softfloat
   3237 * comparison functions float*_{eq,le,lt} all return false for unordered.
   3238 */
   3239#define DO_GE16(X, Y, S) float16_le(Y, X, S)
   3240#define DO_GE32(X, Y, S) float32_le(Y, X, S)
   3241#define DO_GT16(X, Y, S) float16_lt(Y, X, S)
   3242#define DO_GT32(X, Y, S) float32_lt(Y, X, S)
   3243
   3244DO_VCMP_FP_BOTH(vfcmpeqh, vfcmpeq_scalarh, 2, float16, float16_eq)
   3245DO_VCMP_FP_BOTH(vfcmpeqs, vfcmpeq_scalars, 4, float32, float32_eq)
   3246
   3247DO_VCMP_FP_BOTH(vfcmpneh, vfcmpne_scalarh, 2, float16, !float16_eq)
   3248DO_VCMP_FP_BOTH(vfcmpnes, vfcmpne_scalars, 4, float32, !float32_eq)
   3249
   3250DO_VCMP_FP_BOTH(vfcmpgeh, vfcmpge_scalarh, 2, float16, DO_GE16)
   3251DO_VCMP_FP_BOTH(vfcmpges, vfcmpge_scalars, 4, float32, DO_GE32)
   3252
   3253DO_VCMP_FP_BOTH(vfcmplth, vfcmplt_scalarh, 2, float16, !DO_GE16)
   3254DO_VCMP_FP_BOTH(vfcmplts, vfcmplt_scalars, 4, float32, !DO_GE32)
   3255
   3256DO_VCMP_FP_BOTH(vfcmpgth, vfcmpgt_scalarh, 2, float16, DO_GT16)
   3257DO_VCMP_FP_BOTH(vfcmpgts, vfcmpgt_scalars, 4, float32, DO_GT32)
   3258
   3259DO_VCMP_FP_BOTH(vfcmpleh, vfcmple_scalarh, 2, float16, !DO_GT16)
   3260DO_VCMP_FP_BOTH(vfcmples, vfcmple_scalars, 4, float32, !DO_GT32)
   3261
   3262#define DO_VCVT_FIXED(OP, ESIZE, TYPE, FN)                              \
   3263    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vm,   \
   3264                                uint32_t shift)                         \
   3265    {                                                                   \
   3266        TYPE *d = vd, *m = vm;                                          \
   3267        TYPE r;                                                         \
   3268        uint16_t mask = mve_element_mask(env);                          \
   3269        unsigned e;                                                     \
   3270        float_status *fpst;                                             \
   3271        float_status scratch_fpst;                                      \
   3272        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   3273            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   3274                continue;                                               \
   3275            }                                                           \
   3276            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3277                &env->vfp.standard_fp_status;                           \
   3278            if (!(mask & 1)) {                                          \
   3279                /* We need the result but without updating flags */     \
   3280                scratch_fpst = *fpst;                                   \
   3281                fpst = &scratch_fpst;                                   \
   3282            }                                                           \
   3283            r = FN(m[H##ESIZE(e)], shift, fpst);                        \
   3284            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   3285        }                                                               \
   3286        mve_advance_vpt(env);                                           \
   3287    }
   3288
   3289DO_VCVT_FIXED(vcvt_sh, 2, int16_t, helper_vfp_shtoh)
   3290DO_VCVT_FIXED(vcvt_uh, 2, uint16_t, helper_vfp_uhtoh)
   3291DO_VCVT_FIXED(vcvt_hs, 2, int16_t, helper_vfp_toshh_round_to_zero)
   3292DO_VCVT_FIXED(vcvt_hu, 2, uint16_t, helper_vfp_touhh_round_to_zero)
   3293DO_VCVT_FIXED(vcvt_sf, 4, int32_t, helper_vfp_sltos)
   3294DO_VCVT_FIXED(vcvt_uf, 4, uint32_t, helper_vfp_ultos)
   3295DO_VCVT_FIXED(vcvt_fs, 4, int32_t, helper_vfp_tosls_round_to_zero)
   3296DO_VCVT_FIXED(vcvt_fu, 4, uint32_t, helper_vfp_touls_round_to_zero)
   3297
   3298/* VCVT with specified rmode */
   3299#define DO_VCVT_RMODE(OP, ESIZE, TYPE, FN)                              \
   3300    void HELPER(glue(mve_, OP))(CPUARMState *env,                       \
   3301                                void *vd, void *vm, uint32_t rmode)     \
   3302    {                                                                   \
   3303        TYPE *d = vd, *m = vm;                                          \
   3304        TYPE r;                                                         \
   3305        uint16_t mask = mve_element_mask(env);                          \
   3306        unsigned e;                                                     \
   3307        float_status *fpst;                                             \
   3308        float_status scratch_fpst;                                      \
   3309        float_status *base_fpst = (ESIZE == 2) ?                        \
   3310            &env->vfp.standard_fp_status_f16 :                          \
   3311            &env->vfp.standard_fp_status;                               \
   3312        uint32_t prev_rmode = get_float_rounding_mode(base_fpst);       \
   3313        set_float_rounding_mode(rmode, base_fpst);                      \
   3314        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   3315            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   3316                continue;                                               \
   3317            }                                                           \
   3318            fpst = base_fpst;                                           \
   3319            if (!(mask & 1)) {                                          \
   3320                /* We need the result but without updating flags */     \
   3321                scratch_fpst = *fpst;                                   \
   3322                fpst = &scratch_fpst;                                   \
   3323            }                                                           \
   3324            r = FN(m[H##ESIZE(e)], 0, fpst);                            \
   3325            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   3326        }                                                               \
   3327        set_float_rounding_mode(prev_rmode, base_fpst);                 \
   3328        mve_advance_vpt(env);                                           \
   3329    }
   3330
   3331DO_VCVT_RMODE(vcvt_rm_sh, 2, uint16_t, helper_vfp_toshh)
   3332DO_VCVT_RMODE(vcvt_rm_uh, 2, uint16_t, helper_vfp_touhh)
   3333DO_VCVT_RMODE(vcvt_rm_ss, 4, uint32_t, helper_vfp_tosls)
   3334DO_VCVT_RMODE(vcvt_rm_us, 4, uint32_t, helper_vfp_touls)
   3335
   3336#define DO_VRINT_RM_H(M, F, S) helper_rinth(M, S)
   3337#define DO_VRINT_RM_S(M, F, S) helper_rints(M, S)
   3338
   3339DO_VCVT_RMODE(vrint_rm_h, 2, uint16_t, DO_VRINT_RM_H)
   3340DO_VCVT_RMODE(vrint_rm_s, 4, uint32_t, DO_VRINT_RM_S)
   3341
   3342/*
   3343 * VCVT between halfprec and singleprec. As usual for halfprec
   3344 * conversions, FZ16 is ignored and AHP is observed.
   3345 */
   3346static void do_vcvt_sh(CPUARMState *env, void *vd, void *vm, int top)
   3347{
   3348    uint16_t *d = vd;
   3349    uint32_t *m = vm;
   3350    uint16_t r;
   3351    uint16_t mask = mve_element_mask(env);
   3352    bool ieee = !(env->vfp.xregs[ARM_VFP_FPSCR] & FPCR_AHP);
   3353    unsigned e;
   3354    float_status *fpst;
   3355    float_status scratch_fpst;
   3356    float_status *base_fpst = &env->vfp.standard_fp_status;
   3357    bool old_fz = get_flush_to_zero(base_fpst);
   3358    set_flush_to_zero(false, base_fpst);
   3359    for (e = 0; e < 16 / 4; e++, mask >>= 4) {
   3360        if ((mask & MAKE_64BIT_MASK(0, 4)) == 0) {
   3361            continue;
   3362        }
   3363        fpst = base_fpst;
   3364        if (!(mask & 1)) {
   3365            /* We need the result but without updating flags */
   3366            scratch_fpst = *fpst;
   3367            fpst = &scratch_fpst;
   3368        }
   3369        r = float32_to_float16(m[H4(e)], ieee, fpst);
   3370        mergemask(&d[H2(e * 2 + top)], r, mask >> (top * 2));
   3371    }
   3372    set_flush_to_zero(old_fz, base_fpst);
   3373    mve_advance_vpt(env);
   3374}
   3375
   3376static void do_vcvt_hs(CPUARMState *env, void *vd, void *vm, int top)
   3377{
   3378    uint32_t *d = vd;
   3379    uint16_t *m = vm;
   3380    uint32_t r;
   3381    uint16_t mask = mve_element_mask(env);
   3382    bool ieee = !(env->vfp.xregs[ARM_VFP_FPSCR] & FPCR_AHP);
   3383    unsigned e;
   3384    float_status *fpst;
   3385    float_status scratch_fpst;
   3386    float_status *base_fpst = &env->vfp.standard_fp_status;
   3387    bool old_fiz = get_flush_inputs_to_zero(base_fpst);
   3388    set_flush_inputs_to_zero(false, base_fpst);
   3389    for (e = 0; e < 16 / 4; e++, mask >>= 4) {
   3390        if ((mask & MAKE_64BIT_MASK(0, 4)) == 0) {
   3391            continue;
   3392        }
   3393        fpst = base_fpst;
   3394        if (!(mask & (1 << (top * 2)))) {
   3395            /* We need the result but without updating flags */
   3396            scratch_fpst = *fpst;
   3397            fpst = &scratch_fpst;
   3398        }
   3399        r = float16_to_float32(m[H2(e * 2 + top)], ieee, fpst);
   3400        mergemask(&d[H4(e)], r, mask);
   3401    }
   3402    set_flush_inputs_to_zero(old_fiz, base_fpst);
   3403    mve_advance_vpt(env);
   3404}
   3405
   3406void HELPER(mve_vcvtb_sh)(CPUARMState *env, void *vd, void *vm)
   3407{
   3408    do_vcvt_sh(env, vd, vm, 0);
   3409}
   3410void HELPER(mve_vcvtt_sh)(CPUARMState *env, void *vd, void *vm)
   3411{
   3412    do_vcvt_sh(env, vd, vm, 1);
   3413}
   3414void HELPER(mve_vcvtb_hs)(CPUARMState *env, void *vd, void *vm)
   3415{
   3416    do_vcvt_hs(env, vd, vm, 0);
   3417}
   3418void HELPER(mve_vcvtt_hs)(CPUARMState *env, void *vd, void *vm)
   3419{
   3420    do_vcvt_hs(env, vd, vm, 1);
   3421}
   3422
   3423#define DO_1OP_FP(OP, ESIZE, TYPE, FN)                                  \
   3424    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vm)   \
   3425    {                                                                   \
   3426        TYPE *d = vd, *m = vm;                                          \
   3427        TYPE r;                                                         \
   3428        uint16_t mask = mve_element_mask(env);                          \
   3429        unsigned e;                                                     \
   3430        float_status *fpst;                                             \
   3431        float_status scratch_fpst;                                      \
   3432        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
   3433            if ((mask & MAKE_64BIT_MASK(0, ESIZE)) == 0) {              \
   3434                continue;                                               \
   3435            }                                                           \
   3436            fpst = (ESIZE == 2) ? &env->vfp.standard_fp_status_f16 :    \
   3437                &env->vfp.standard_fp_status;                           \
   3438            if (!(mask & 1)) {                                          \
   3439                /* We need the result but without updating flags */     \
   3440                scratch_fpst = *fpst;                                   \
   3441                fpst = &scratch_fpst;                                   \
   3442            }                                                           \
   3443            r = FN(m[H##ESIZE(e)], fpst);                               \
   3444            mergemask(&d[H##ESIZE(e)], r, mask);                        \
   3445        }                                                               \
   3446        mve_advance_vpt(env);                                           \
   3447    }
   3448
   3449DO_1OP_FP(vrintx_h, 2, float16, float16_round_to_int)
   3450DO_1OP_FP(vrintx_s, 4, float32, float32_round_to_int)