target/arm/vec_internal.h

   1 /*
   2  * ARM AdvSIMD / SVE Vector Helpers
   3  *
   4  * Copyright (c) 2020 Linaro
   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 #ifndef TARGET_ARM_VEC_INTERNALS_H
  21 #define TARGET_ARM_VEC_INTERNALS_H
  22
  23 /*
  24  * Note that vector data is stored in host-endian 64-bit chunks,
  25  * so addressing units smaller than that needs a host-endian fixup.
  26  *
  27  * The H<N> macros are used when indexing an array of elements of size N.
  28  *
  29  * The H1_<N> macros are used when performing byte arithmetic and then
  30  * casting the final pointer to a type of size N.
  31  */
  32 #ifdef HOST_WORDS_BIGENDIAN
  33 #define H1(x)   ((x) ^ 7)
  34 #define H1_2(x) ((x) ^ 6)
  35 #define H1_4(x) ((x) ^ 4)
  36 #define H2(x)   ((x) ^ 3)
  37 #define H4(x)   ((x) ^ 1)
  38 #else
  39 #define H1(x)   (x)
  40 #define H1_2(x) (x)
  41 #define H1_4(x) (x)
  42 #define H2(x)   (x)
  43 #define H4(x)   (x)
  44 #endif
  45 /*
  46  * Access to 64-bit elements isn't host-endian dependent; we provide H8
  47  * and H1_8 so that when a function is being generated from a macro we
  48  * can pass these rather than an empty macro argument, for clarity.
  49  */
  50 #define H8(x)   (x)
  51 #define H1_8(x) (x)
  52
  53 static inline void clear_tail(void *vd, uintptr_t opr_sz, uintptr_t max_sz)
  54 {
  55     uint64_t *d = vd + opr_sz;
  56     uintptr_t i;
  57
  58     for (i = opr_sz; i < max_sz; i += 8) {
  59         *d++ = 0;
  60     }
  61 }
  62
  63 static inline int32_t do_sqrshl_bhs(int32_t src, int32_t shift, int bits,
  64                                     bool round, uint32_t *sat)
  65 {
  66     if (shift <= -bits) {
  67         /* Rounding the sign bit always produces 0. */
  68         if (round) {
  69             return 0;
  70         }
  71         return src >> 31;
  72     } else if (shift < 0) {
  73         if (round) {
  74             src >>= -shift - 1;
  75             return (src >> 1) + (src & 1);
  76         }
  77         return src >> -shift;
  78     } else if (shift < bits) {
  79         int32_t val = src << shift;
  80         if (bits == 32) {
  81             if (!sat || val >> shift == src) {
  82                 return val;
  83             }
  84         } else {
  85             int32_t extval = sextract32(val, 0, bits);
  86             if (!sat || val == extval) {
  87                 return extval;
  88             }
  89         }
  90     } else if (!sat || src == 0) {
  91         return 0;
  92     }
  93
  94     *sat = 1;
  95     return (1u << (bits - 1)) - (src >= 0);
  96 }
  97
  98 static inline uint32_t do_uqrshl_bhs(uint32_t src, int32_t shift, int bits,
  99                                      bool round, uint32_t *sat)
 100 {
 101     if (shift <= -(bits + round)) {
 102         return 0;
 103     } else if (shift < 0) {
 104         if (round) {
 105             src >>= -shift - 1;
 106             return (src >> 1) + (src & 1);
 107         }
 108         return src >> -shift;
 109     } else if (shift < bits) {
 110         uint32_t val = src << shift;
 111         if (bits == 32) {
 112             if (!sat || val >> shift == src) {
 113                 return val;
 114             }
 115         } else {
 116             uint32_t extval = extract32(val, 0, bits);
 117             if (!sat || val == extval) {
 118                 return extval;
 119             }
 120         }
 121     } else if (!sat || src == 0) {
 122         return 0;
 123     }
 124
 125     *sat = 1;
 126     return MAKE_64BIT_MASK(0, bits);
 127 }
 128
 129 static inline int32_t do_suqrshl_bhs(int32_t src, int32_t shift, int bits,
 130                                      bool round, uint32_t *sat)
 131 {
 132     if (sat && src < 0) {
 133         *sat = 1;
 134         return 0;
 135     }
 136     return do_uqrshl_bhs(src, shift, bits, round, sat);
 137 }
 138
 139 static inline int64_t do_sqrshl_d(int64_t src, int64_t shift,
 140                                   bool round, uint32_t *sat)
 141 {
 142     if (shift <= -64) {
 143         /* Rounding the sign bit always produces 0. */
 144         if (round) {
 145             return 0;
 146         }
 147         return src >> 63;
 148     } else if (shift < 0) {
 149         if (round) {
 150             src >>= -shift - 1;
 151             return (src >> 1) + (src & 1);
 152         }
 153         return src >> -shift;
 154     } else if (shift < 64) {
 155         int64_t val = src << shift;
 156         if (!sat || val >> shift == src) {
 157             return val;
 158         }
 159     } else if (!sat || src == 0) {
 160         return 0;
 161     }
 162
 163     *sat = 1;
 164     return src < 0 ? INT64_MIN : INT64_MAX;
 165 }
 166
 167 static inline uint64_t do_uqrshl_d(uint64_t src, int64_t shift,
 168                                    bool round, uint32_t *sat)
 169 {
 170     if (shift <= -(64 + round)) {
 171         return 0;
 172     } else if (shift < 0) {
 173         if (round) {
 174             src >>= -shift - 1;
 175             return (src >> 1) + (src & 1);
 176         }
 177         return src >> -shift;
 178     } else if (shift < 64) {
 179         uint64_t val = src << shift;
 180         if (!sat || val >> shift == src) {
 181             return val;
 182         }
 183     } else if (!sat || src == 0) {
 184         return 0;
 185     }
 186
 187     *sat = 1;
 188     return UINT64_MAX;
 189 }
 190
 191 static inline int64_t do_suqrshl_d(int64_t src, int64_t shift,
 192                                    bool round, uint32_t *sat)
 193 {
 194     if (sat && src < 0) {
 195         *sat = 1;
 196         return 0;
 197     }
 198     return do_uqrshl_d(src, shift, round, sat);
 199 }
 200
 201 int8_t do_sqrdmlah_b(int8_t, int8_t, int8_t, bool, bool);
 202 int16_t do_sqrdmlah_h(int16_t, int16_t, int16_t, bool, bool, uint32_t *);
 203 int32_t do_sqrdmlah_s(int32_t, int32_t, int32_t, bool, bool, uint32_t *);
 204 int64_t do_sqrdmlah_d(int64_t, int64_t, int64_t, bool, bool);
 205
 206 #endif /* TARGET_ARM_VEC_INTERNALS_H */