| 1 | const builtin = @import("builtin"); |
| 2 | |
| 3 | const std = @import("std"); |
| 4 | const Log2Int = std.math.Log2Int; |
| 5 | |
| 6 | const compiler_rt = @import("../compiler_rt.zig"); |
| 7 | const symbol = compiler_rt.symbol; |
| 8 | const HalveInt = compiler_rt.HalveInt; |
| 9 | |
| 10 | comptime { |
| 11 | symbol(&__umodti3, "__umodti3"); |
| 12 | symbol(&__modti3, "__modti3"); |
| 13 | symbol(&__udivti3, "__udivti3"); |
| 14 | symbol(&__divti3, "__divti3"); |
| 15 | symbol(&__udivmodti4, "__udivmodti4"); |
| 16 | } |
| 17 | |
| 18 | pub fn __udivmodti4(a: u128, b: u128, maybe_rem: ?*u128) callconv(.c) u128 { |
| 19 | return udivmod(u128, a, b, maybe_rem); |
| 20 | } |
| 21 | |
| 22 | pub fn __divti3(a: i128, b: i128) callconv(.c) i128 { |
| 23 | return div(a, b); |
| 24 | } |
| 25 | |
| 26 | inline fn div(a: i128, b: i128) i128 { |
| 27 | const s_a = a >> (128 - 1); |
| 28 | const s_b = b >> (128 - 1); |
| 29 | |
| 30 | const an = (a ^ s_a) -% s_a; |
| 31 | const bn = (b ^ s_b) -% s_b; |
| 32 | |
| 33 | const r = udivmod(u128, @bitCast(an), @bitCast(bn), null); |
| 34 | const s = s_a ^ s_b; |
| 35 | return (@as(i128, @bitCast(r)) ^ s) -% s; |
| 36 | } |
| 37 | |
| 38 | pub fn __udivti3(a: u128, b: u128) callconv(.c) u128 { |
| 39 | return udivmod(u128, a, b, null); |
| 40 | } |
| 41 | |
| 42 | pub fn __modti3(a: i128, b: i128) callconv(.c) i128 { |
| 43 | return mod(a, b); |
| 44 | } |
| 45 | |
| 46 | inline fn mod(a: i128, b: i128) i128 { |
| 47 | const s_a = a >> (128 - 1); // s = a < 0 ? -1 : 0 |
| 48 | const s_b = b >> (128 - 1); // s = b < 0 ? -1 : 0 |
| 49 | |
| 50 | const an = (a ^ s_a) -% s_a; // negate if s == -1 |
| 51 | const bn = (b ^ s_b) -% s_b; // negate if s == -1 |
| 52 | |
| 53 | var r: u128 = undefined; |
| 54 | _ = udivmod(u128, @as(u128, @bitCast(an)), @as(u128, @bitCast(bn)), &r); |
| 55 | return (@as(i128, @bitCast(r)) ^ s_a) -% s_a; // negate if s == -1 |
| 56 | } |
| 57 | |
| 58 | pub fn __umodti3(a: u128, b: u128) callconv(.c) u128 { |
| 59 | var r: u128 = undefined; |
| 60 | _ = udivmod(u128, a, b, &r); |
| 61 | return r; |
| 62 | } |
| 63 | |
| 64 | // Let _u1 and _u0 be the high and low limbs of U respectively. |
| 65 | // Returns U / v_ and sets r = U % v_. |
| 66 | fn divwide_generic(comptime T: type, _u1: T, _u0: T, v_: T, r: *T) T { |
| 67 | const HalfT = HalveInt(T, false).HalfT; |
| 68 | @setRuntimeSafety(compiler_rt.test_safety); |
| 69 | var v = v_; |
| 70 | |
| 71 | const b = @as(T, 1) << (@bitSizeOf(T) / 2); |
| 72 | var un64: T = undefined; |
| 73 | var un10: T = undefined; |
| 74 | |
| 75 | const s: Log2Int(T) = @intCast(@clz(v)); |
| 76 | if (s > 0) { |
| 77 | // Normalize divisor |
| 78 | v <<= s; |
| 79 | un64 = (_u1 << s) | (_u0 >> @intCast((@bitSizeOf(T) - @as(T, @intCast(s))))); |
| 80 | un10 = _u0 << s; |
| 81 | } else { |
| 82 | // Avoid undefined behavior of (u0 >> @bitSizeOf(T)) |
| 83 | un64 = _u1; |
| 84 | un10 = _u0; |
| 85 | } |
| 86 | |
| 87 | // Break divisor up into two 32-bit digits |
| 88 | const vn1 = v >> (@bitSizeOf(T) / 2); |
| 89 | const vn0 = v & std.math.maxInt(HalfT); |
| 90 | |
| 91 | // Break right half of dividend into two digits |
| 92 | const un1 = un10 >> (@bitSizeOf(T) / 2); |
| 93 | const un0 = un10 & std.math.maxInt(HalfT); |
| 94 | |
| 95 | // Compute the first quotient digit, q1 |
| 96 | var q1 = un64 / vn1; |
| 97 | var rhat = un64 -% q1 *% vn1; |
| 98 | |
| 99 | // q1 has at most error 2. No more than 2 iterations |
| 100 | while (q1 >= b or q1 * vn0 > b * rhat + un1) { |
| 101 | q1 -= 1; |
| 102 | rhat += vn1; |
| 103 | if (rhat >= b) break; |
| 104 | } |
| 105 | |
| 106 | const un21 = un64 *% b +% un1 -% q1 *% v; |
| 107 | |
| 108 | // Compute the second quotient digit |
| 109 | var q0 = un21 / vn1; |
| 110 | rhat = un21 -% q0 *% vn1; |
| 111 | |
| 112 | // q0 has at most error 2. No more than 2 iterations. |
| 113 | while (q0 >= b or q0 * vn0 > b * rhat + un0) { |
| 114 | q0 -= 1; |
| 115 | rhat += vn1; |
| 116 | if (rhat >= b) break; |
| 117 | } |
| 118 | |
| 119 | r.* = (un21 *% b +% un0 -% q0 *% v) >> s; |
| 120 | return q1 *% b +% q0; |
| 121 | } |
| 122 | |
| 123 | fn divwide(comptime T: type, _u1: T, _u0: T, v: T, r: *T) T { |
| 124 | @setRuntimeSafety(compiler_rt.test_safety); |
| 125 | if (T == u64 and builtin.target.cpu.arch == .x86_64 and builtin.target.os.tag != .windows) { |
| 126 | var rem: T = undefined; |
| 127 | const quo = asm ( |
| 128 | \\divq %[v] |
| 129 | : [_] "={rax}" (-> T), |
| 130 | [_] "={rdx}" (rem), |
| 131 | : [v] "r" (v), |
| 132 | [_] "{rax}" (_u0), |
| 133 | [_] "{rdx}" (_u1), |
| 134 | ); |
| 135 | r.* = rem; |
| 136 | return quo; |
| 137 | } else { |
| 138 | return divwide_generic(T, _u1, _u0, v, r); |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | // Returns a_ / b_ and sets maybe_rem = a_ % b. |
| 143 | pub fn udivmod(comptime T: type, a_: T, b_: T, maybe_rem: ?*T) T { |
| 144 | @setRuntimeSafety(compiler_rt.test_safety); |
| 145 | const HalfT = HalveInt(T, false).HalfT; |
| 146 | const half_bits = @bitSizeOf(HalfT); |
| 147 | |
| 148 | if (b_ > a_) { |
| 149 | if (maybe_rem) |rem| { |
| 150 | rem.* = a_; |
| 151 | } |
| 152 | return 0; |
| 153 | } |
| 154 | |
| 155 | const a: [2]HalfT = @bitCast(a_); // [0] is low bits, [1] is high bits |
| 156 | const b: [2]HalfT = @bitCast(b_); // [0] is low bits, [1] is high bits |
| 157 | var q: [2]HalfT = undefined; |
| 158 | var r: [2]HalfT = undefined; |
| 159 | |
| 160 | // When the divisor fits in 64 bits, we can use an optimized path |
| 161 | if (b[1] == 0) { |
| 162 | r[1] = 0; |
| 163 | if (a[1] < b[0]) { |
| 164 | // The result fits in 64 bits |
| 165 | q[1] = 0; |
| 166 | q[0] = divwide(HalfT, a[1], a[0], b[0], &r[0]); |
| 167 | } else { |
| 168 | // First, divide with the high part to get the remainder. After that a_hi < b_lo. |
| 169 | q[1] = a[1] / b[0]; |
| 170 | q[0] = divwide(HalfT, a[1] % b[0], a[0], b[0], &r[0]); |
| 171 | } |
| 172 | if (maybe_rem) |rem| { |
| 173 | rem.* = @bitCast(r); |
| 174 | } |
| 175 | return @bitCast(q); |
| 176 | } |
| 177 | |
| 178 | // Large-divisor case: b[1] != 0, so the quotient fits in one HalfT word. |
| 179 | // |
| 180 | // Trial quotient via divwide (Knuth Vol 2, Section 4.3.1): |
| 181 | // Normalize the divisor so its high half has the MSB set, then use divwide |
| 182 | // on the top bits to get a trial quotient that is at most 1 too large. |
| 183 | // This replaces the O(shift) bit-by-bit loop with O(1) operations. |
| 184 | const s: Log2Int(HalfT) = @intCast(@clz(b[1])); |
| 185 | |
| 186 | if (s == 0) { |
| 187 | // b[1] already has its MSB set, so b >= 2^(T_bits - 1). Since a >= b |
| 188 | // (we passed the b_ > a_ check), a >= 2^(T_bits - 1) too, meaning |
| 189 | // a[1] also has its MSB set. Therefore a / b < 2, and the quotient |
| 190 | // is exactly 1. |
| 191 | q = @bitCast(@as(T, 0)); |
| 192 | q[0] = 1; |
| 193 | if (maybe_rem) |rem| { |
| 194 | rem.* = a_ - b_; |
| 195 | } |
| 196 | return @bitCast(q); |
| 197 | } |
| 198 | |
| 199 | // Normalize b: shift left by s so bn_hi has its MSB set. |
| 200 | const sr: Log2Int(HalfT) = @intCast(half_bits - @as( |
| 201 | std.math.IntFittingRange(0, half_bits), |
| 202 | @intCast(s), |
| 203 | )); |
| 204 | const bn_hi: HalfT = (b[1] << s) | (b[0] >> sr); |
| 205 | |
| 206 | // Trial numerator: the top (half_bits + s) bits of (a << s), as [a2:a1]. |
| 207 | // a2 < bn_hi is guaranteed since a2 < 2^s and bn_hi >= 2^(half_bits - 1). |
| 208 | const a2: HalfT = a[1] >> sr; |
| 209 | const a1: HalfT = (a[1] << s) | (a[0] >> sr); |
| 210 | |
| 211 | // Trial quotient via divwide: q_hat = floor([a2:a1] / bn_hi). |
| 212 | // By Knuth's theorem (normalized divisor), q <= q_hat <= q + 1. |
| 213 | var r_tmp: HalfT = undefined; |
| 214 | var q_hat: HalfT = divwide(HalfT, a2, a1, bn_hi, &r_tmp); |
| 215 | |
| 216 | // Verify: q_hat * b must not exceed a. |
| 217 | // Compute the product using HalfT * HalfT -> T widening multiplications, |
| 218 | // which are native single-instruction ops when HalfT fits in a register |
| 219 | // (e.g. u64 * u64 -> u128 via mulq on x86_64, mul on aarch64). |
| 220 | // product = q_hat * [b[1]:b[0]] = [p_top : p_mid : p_lo] (3 half-words) |
| 221 | const prod_lo: T = @as(T, q_hat) * @as(T, b[0]); |
| 222 | const prod_hi: T = @as(T, q_hat) * @as(T, b[1]); |
| 223 | |
| 224 | const prod_lo_parts: [2]HalfT = @bitCast(prod_lo); |
| 225 | const prod_hi_parts: [2]HalfT = @bitCast(prod_hi); |
| 226 | |
| 227 | const mid_add = @addWithOverflow(prod_hi_parts[0], prod_lo_parts[1]); |
| 228 | var p_mid: HalfT = mid_add[0]; |
| 229 | const p_top: HalfT = prod_hi_parts[1] +% @as(HalfT, mid_add[1]); |
| 230 | var p_lo: HalfT = prod_lo_parts[0]; |
| 231 | |
| 232 | // If product > a, decrement q_hat (at most once, guaranteed by Knuth). |
| 233 | if (p_top > 0 or p_mid > a[1] or (p_mid == a[1] and p_lo > a[0])) { |
| 234 | q_hat -= 1; |
| 235 | // Subtract b from the product for correct remainder computation. |
| 236 | // After correction, (q_hat * b) fits in T bits, so borrows into |
| 237 | // p_top cancel it to zero -- we only need [p_mid:p_lo]. |
| 238 | const sub_lo = @subWithOverflow(p_lo, b[0]); |
| 239 | p_lo = sub_lo[0]; |
| 240 | const sub_mid = @subWithOverflow(p_mid, b[1]); |
| 241 | const sub_mid2 = @subWithOverflow(sub_mid[0], @as(HalfT, sub_lo[1])); |
| 242 | p_mid = sub_mid2[0]; |
| 243 | } |
| 244 | |
| 245 | q = @bitCast(@as(T, 0)); |
| 246 | q[0] = q_hat; |
| 247 | |
| 248 | if (maybe_rem) |rem| { |
| 249 | // remainder = a - q_hat * b = [a[1]:a[0]] - [p_mid:p_lo] |
| 250 | // This subtraction is non-negative since q_hat <= true quotient. |
| 251 | const rem_lo = @subWithOverflow(a[0], p_lo); |
| 252 | r[0] = rem_lo[0]; |
| 253 | const rem_hi = @subWithOverflow(a[1], p_mid); |
| 254 | const rem_hi2 = @subWithOverflow(rem_hi[0], @as(HalfT, rem_lo[1])); |
| 255 | r[1] = rem_hi2[0]; |
| 256 | rem.* = @bitCast(r); |
| 257 | } |
| 258 | return @bitCast(q); |
| 259 | } |
| 260 | |
| 261 | test { |
| 262 | _ = @import("modti3_test.zig"); |
| 263 | _ = @import("divti3_test.zig"); |
| 264 | _ = @import("udivmodti4_test.zig"); |
| 265 | } |