25    input data by 64x64-bit multiplication and mixing hi- and low-parts
28    with the equal probability of their bit values for the
38    strings (at least up to 512-bit) on Haswell and Power7.  The MUM bulk
59 /* Macro saying to use 128-bit integers implemented by GCC for some
90    probability of their bit values.  They are used to randomize input
107 /* Multiply 64-bit V and P and return sum of high and low parts of the
110 _mum (uint64_t v, uint64_t p) {  in _mum()  argument
114   /* AARCH64 needs 2 insns to calculate 128-bit result of the  in _mum()
116      function doing 128x128->128 bit multiplication.  The function is  in _mum()
118   lo = v * p, hi;  in _mum()
119   asm ("umulh %0, %1, %2" : "=r" (hi) : "r" (v), "r" (p));  in _mum()
121   __uint128_t r = (__uint128_t) v * (__uint128_t) p;  in _mum()
126   /* Implementation of 64x64->128-bit multiplication by four 32x32->64  in _mum()
127      bit multiplication.  */  in _mum()
128   uint64_t hv = v >> 32, hp = p >> 32;  in _mum()
129   uint64_t lv = (uint32_t) v, lp = (uint32_t) p;  in _mum()
136   /* We could ignore a carry bit here if we did not care about the  in _mum()
137      same hash for 32-bit and 64-bit targets.  */  in _mum()
171 _mum_le (uint64_t v) {  in _mum_le()  argument
173   return v;  in _mum_le()
175   return _mum_bswap64 (v);  in _mum_le()
182 _mum_le32 (uint32_t v) {  in _mum_le32()  argument
184   return v;  in _mum_le32()
186   return _mum_bswap32 (v);  in _mum_le32()
197    The macro value affects the result hash for strings > 128 bit.  The
218 static inline uint64_t _MUM_OPTIMIZE("unroll-loops")
229        64x64->128-bit multiplication.  AVX2 currently only have a  in _mum_hash_aligned()
230        vector insn for 4 32x32->64-bit multiplication.  */  in _mum_hash_aligned()
233     len -= _MUM_UNROLL_FACTOR * sizeof (uint64_t);  in _mum_hash_aligned()
235     /* We will use the same prime numbers on the next iterations --  in _mum_hash_aligned()
242   len -= n * sizeof (uint64_t); str += n * sizeof (uint64_t);  in _mum_hash_aligned()
288 /* We want to use AVX2 insn MULX instead of generic x86-64 MULQ where
290    3-cycles vs. 4 for MULX, MULX permits more freedom in insn
322 _MUM_TARGET("inline-all-stringops")
338       len -= block_len;  in _mum_hash_default()
411     avx2_support =  __builtin_cpu_supports ("avx2") ? 1 : -1;  in mum_hash()