| 1 | //===-- udivmodti4.c - Implement __udivmodti4 -----------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements __udivmodti4 for the compiler_rt library. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "int_lib.h" |
| 14 | |
| 15 | #ifdef CRT_HAS_128BIT |
| 16 | |
| 17 | // Returns the 128 bit division result by 64 bit. Result must fit in 64 bits. |
| 18 | // Remainder stored in r. |
| 19 | // Taken and adjusted from libdivide libdivide_128_div_64_to_64 division |
| 20 | // fallback. For a correctness proof see the reference for this algorithm |
| 21 | // in Knuth, Volume 2, section 4.3.1, Algorithm D. |
| 22 | UNUSED |
| 23 | static inline du_int udiv128by64to64default(du_int u1, du_int u0, du_int v, |
| 24 | du_int *r) { |
| 25 | const unsigned n_udword_bits = sizeof(du_int) * CHAR_BIT; |
| 26 | const du_int b = (1ULL << (n_udword_bits / 2)); // Number base (32 bits) |
| 27 | du_int un1, un0; // Norm. dividend LSD's |
| 28 | du_int vn1, vn0; // Norm. divisor digits |
| 29 | du_int q1, q0; // Quotient digits |
| 30 | du_int un64, un21, un10; // Dividend digit pairs |
| 31 | du_int rhat; // A remainder |
| 32 | si_int s; // Shift amount for normalization |
| 33 | |
| 34 | s = __builtin_clzll(v); |
| 35 | if (s > 0) { |
| 36 | // Normalize the divisor. |
| 37 | v = v << s; |
| 38 | un64 = (u1 << s) | (u0 >> (n_udword_bits - s)); |
| 39 | un10 = u0 << s; // Shift dividend left |
| 40 | } else { |
| 41 | // Avoid undefined behavior of (u0 >> 64). |
| 42 | un64 = u1; |
| 43 | un10 = u0; |
| 44 | } |
| 45 | |
| 46 | // Break divisor up into two 32-bit digits. |
| 47 | vn1 = v >> (n_udword_bits / 2); |
| 48 | vn0 = v & 0xFFFFFFFF; |
| 49 | |
| 50 | // Break right half of dividend into two digits. |
| 51 | un1 = un10 >> (n_udword_bits / 2); |
| 52 | un0 = un10 & 0xFFFFFFFF; |
| 53 | |
| 54 | // Compute the first quotient digit, q1. |
| 55 | q1 = un64 / vn1; |
| 56 | rhat = un64 - q1 * vn1; |
| 57 | |
| 58 | // q1 has at most error 2. No more than 2 iterations. |
| 59 | while (q1 >= b || q1 * vn0 > b * rhat + un1) { |
| 60 | q1 = q1 - 1; |
| 61 | rhat = rhat + vn1; |
| 62 | if (rhat >= b) |
| 63 | break; |
| 64 | } |
| 65 | |
| 66 | un21 = un64 * b + un1 - q1 * v; |
| 67 | |
| 68 | // Compute the second quotient digit. |
| 69 | q0 = un21 / vn1; |
| 70 | rhat = un21 - q0 * vn1; |
| 71 | |
| 72 | // q0 has at most error 2. No more than 2 iterations. |
| 73 | while (q0 >= b || q0 * vn0 > b * rhat + un0) { |
| 74 | q0 = q0 - 1; |
| 75 | rhat = rhat + vn1; |
| 76 | if (rhat >= b) |
| 77 | break; |
| 78 | } |
| 79 | |
| 80 | *r = (un21 * b + un0 - q0 * v) >> s; |
| 81 | return q1 * b + q0; |
| 82 | } |
| 83 | |
| 84 | static inline du_int udiv128by64to64(du_int u1, du_int u0, du_int v, |
| 85 | du_int *r) { |
| 86 | #if defined(__x86_64__) && !defined(__arm64ec__) |
| 87 | du_int result; |
| 88 | __asm__("divq %[v]" |
| 89 | : "=a" (result), "=d" (*r) |
| 90 | : [ v ] "r" (v), "a" (u0), "d" (u1)); |
| 91 | return result; |
| 92 | #else |
| 93 | return udiv128by64to64default(u1, u0, v, r); |
| 94 | #endif |
| 95 | } |
| 96 | |
| 97 | // Effects: if rem != 0, *rem = a % b |
| 98 | // Returns: a / b |
| 99 | |
| 100 | COMPILER_RT_ABI tu_int __udivmodti4(tu_int a, tu_int b, tu_int *rem) { |
| 101 | const unsigned n_utword_bits = sizeof(tu_int) * CHAR_BIT; |
| 102 | utwords dividend; |
| 103 | dividend.all = a; |
| 104 | utwords divisor; |
| 105 | divisor.all = b; |
| 106 | utwords quotient; |
| 107 | utwords remainder; |
| 108 | if (divisor.all > dividend.all) { |
| 109 | if (rem) |
| 110 | *rem = dividend.all; |
| 111 | return 0; |
| 112 | } |
| 113 | // When the divisor fits in 64 bits, we can use an optimized path. |
| 114 | if (divisor.s.high == 0) { |
| 115 | remainder.s.high = 0; |
| 116 | if (dividend.s.high < divisor.s.low) { |
| 117 | // The result fits in 64 bits. |
| 118 | quotient.s.low = udiv128by64to64(u1: dividend.s.high, u0: dividend.s.low, |
| 119 | v: divisor.s.low, r: &remainder.s.low); |
| 120 | quotient.s.high = 0; |
| 121 | } else { |
| 122 | // First, divide with the high part to get the remainder in dividend.s.high. |
| 123 | // After that dividend.s.high < divisor.s.low. |
| 124 | quotient.s.high = dividend.s.high / divisor.s.low; |
| 125 | dividend.s.high = dividend.s.high % divisor.s.low; |
| 126 | quotient.s.low = udiv128by64to64(u1: dividend.s.high, u0: dividend.s.low, |
| 127 | v: divisor.s.low, r: &remainder.s.low); |
| 128 | } |
| 129 | if (rem) |
| 130 | *rem = remainder.all; |
| 131 | return quotient.all; |
| 132 | } |
| 133 | // 0 <= shift <= 63. |
| 134 | si_int shift = |
| 135 | __builtin_clzll(divisor.s.high) - __builtin_clzll(dividend.s.high); |
| 136 | divisor.all <<= shift; |
| 137 | quotient.s.high = 0; |
| 138 | quotient.s.low = 0; |
| 139 | for (; shift >= 0; --shift) { |
| 140 | quotient.s.low <<= 1; |
| 141 | // Branch free version of. |
| 142 | // if (dividend.all >= divisor.all) |
| 143 | // { |
| 144 | // dividend.all -= divisor.all; |
| 145 | // carry = 1; |
| 146 | // } |
| 147 | const ti_int s = |
| 148 | (ti_int)(divisor.all - dividend.all - 1) >> (n_utword_bits - 1); |
| 149 | quotient.s.low |= s & 1; |
| 150 | dividend.all -= divisor.all & s; |
| 151 | divisor.all >>= 1; |
| 152 | } |
| 153 | if (rem) |
| 154 | *rem = dividend.all; |
| 155 | return quotient.all; |
| 156 | } |
| 157 | |
| 158 | #endif // CRT_HAS_128BIT |
| 159 | |