#804533 (C) No.3030 ミラー・ラビン素数判定法のテスト

提出ソース
結果

問題	No.3030 ミラー・ラビン素数判定法のテスト
ユーザー	Jashinchan
提出日時	2022-10-06 17:22:41
言語	C (gcc 12.3.0)
結果	AC
実行時間	63 ms / 9,973 ms
コード長	28,733 bytes
コンパイル時間	1,516 ms
コンパイル使用メモリ	62,256 KB
実行使用メモリ	5,376 KB
最終ジャッジ日時	2024-06-11 04:07:36
合計ジャッジ時間	2,095 ms
ジャッジサーバーID （参考情報）	judge4 / judge5
このコードへのチャレンジ
（要ログイン）
テストケース

テストケース表示
入力	結果	実行時間実行使用メモリ
testcase_00	AC	1 ms 5,248 KB
testcase_01	AC	1 ms 5,376 KB
testcase_02	AC	1 ms 5,376 KB
testcase_03	AC	1 ms 5,376 KB
testcase_04	AC	36 ms 5,376 KB
testcase_05	AC	35 ms 5,376 KB
testcase_06	AC	19 ms 5,376 KB
testcase_07	AC	19 ms 5,376 KB
testcase_08	AC	19 ms 5,376 KB
testcase_09	AC	63 ms 5,376 KB
権限があれば一括ダウンロードができます
ソースコード

raw source code
#pragma GCC optimize("O3")
#pragma GCC target("avx2")
#pragma GCC optimize("fast-math")
#pragma GCC optimize("unroll-loops")

#define _GNU_SOURCE
#include <assert.h>
#include <inttypes.h>
#include <limits.h>
#include <math.h>
#include <stdbool.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

typedef   int8_t      i8;
typedef   int16_t     i16;
typedef   int32_t     i32;
typedef   int64_t     i64;
typedef __int128_t    i128;
typedef   uint8_t     u8;
typedef   uint16_t    u16;
typedef   uint32_t    u32;
typedef   uint64_t    u64;
typedef __uint128_t   u128;
typedef   float       f32;
typedef   double      f64;
typedef   long double f80;

#define PRIVATE static
#define PUBLIC

#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define SWAP(a, b)          \
            do {            \
                (a) ^= (b); \
                (b) ^= (a); \
                (a) ^= (b); \
            }               \
            while (0);
#define CTZ32(a)       ((a) ? __builtin_ctz((a)) : (32))
#define CTZ64(a)       ((a) ? __builtin_ctzll((a)) : (64))
#define CLZ32(a)       ((a) ? __builtin_clz((a)) : (32))
#define CLZ64(a)       ((a) ? __builtin_clzll((a)) : (64))
#define POPCNT32(a)    ((a) ? __builtin_popcount((a)) : (0))
#define POPCNT64(a)    ((a) ? __builtin_popcountll((a)) : (0))
#define BIT_WIDTH32(a) ((32) - CLZ32((a)))
#define BIT_WIDTH64(a) ((64) - CLZ64((a)))
#define BIT_FLOOR32(a) ((a) ? ((1u) << (BIT_WIDTH32((a)) - (1))) : (0))
#define BIT_FLOOR64(a) ((a) ? ((1ul) << (BIT_WIDTH64((a)) - (1))) : (0))
#define BIT_CEIL32(a)  (((a) <= 1) ? (1u) : ((1u) << BIT_WIDTH32((a) - (1))))
#define BIT_CEIL64(a)  (((a) <= 1) ? (1ul) : ((1ul) << BIT_WIDTH64((a) - (1))))
#define LSBit(a)       ((a) & (-(a)))
#define CLSBit(a)      ((a) & ((a) - (1)))
#define HAS_SINGLE_BIT32(a) (((a) != (0)) && (CLSBit((a)) == (0)))
#define HAS_SINGLE_BIT64(a) (((a) != (0)) && (CLSBit((a)) == (0)))
    #define _ROTL32_INNER(x, l) (((x) << (l)) | ((x) >> ((-l) & (31))))
    #define _ROTR32_INNER(x, r) (((x) >> (r)) | ((x) << ((-r) & (31))))
    #define _ROTL64_INNER(x, l) (((x) << (l)) | ((x) >> ((-l) & (63))))
    #define _ROTR64_INNER(x, r) (((x) >> (r)) | ((x) << ((-r) & (63))))
#define ROTR32(x, r) (((r) < (0)) ? (_ROTL32_INNER((x), ((u64)(-r) % (32)))) : (_ROTR32_INNER((x), ((r) % (32)))))
#define ROTL32(x, l) ROTR32((x), (-l))
#define ROTR64(x, r) (((r) < (0)) ? (_ROTL64_INNER((x), ((u64)(-r) % (64)))) : (_ROTR64_INNER((x), ((r) % (64)))))
#define ROTL64(x, l) ROTR64((x), (-l))

i32 in_i32(void)
{
    i32 c, x = 0, f = 1;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        if (c == 45)
            f = -f;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return f * x;
}
i64 in_i64(void)
{
    i64 c, x = 0, f = 1;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        if (c == 45)
            f = -f;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return f * x;
}
i128 in_i128(void)
{
    i128 c, x = 0, f = 1;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        if (c == 45)
            f = -f;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return f * x;
}
u32 in_u32(void)
{
    u32 c, x = 0;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        ;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return x;
}
u64 in_u64(void)
{
    u64 c, x = 0;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        ;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return x;
}
u128 in_u128(void)
{
    u128 c, x = 0;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        ;
    while (47 < c && c < 58)
    {
        x = x * 10 + c - 48;
        c = getchar_unlocked();
    }
    return x;
}
static inline void out_i32_inner(i32 x)
{
    if (x >= 10)
        out_i32_inner(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void out_i32(i32 x)
{
    if (x < 0)
    {
        putchar_unlocked('-');
        x = -x;
    }
    out_i32_inner(x);
}
static inline void out_i64_inner(i64 x)
{
    if (x >= 10)
        out_i64_inner(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void out_i64(i64 x)
{
    if (x < 0)
    {
        putchar_unlocked('-');
        x = -x;
    }
    out_i64_inner(x);
}
static inline void out_i128_inner(i128 x)
{
    if (x >= 10)
        out_i128_inner(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void out_i128(i128 x)
{
    if (x < 0)
    {
        putchar_unlocked('-');
        x = -x;
    }
    out_i128_inner(x);
}
void out_u32(u32 x)
{
    if (x >= 10)
        out_u32(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void out_u64(u64 x)
{
    if (x >= 10)
        out_u64(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void out_u128(u128 x)
{
    if (x >= 10)
        out_u128(x / 10);
    putchar_unlocked(x - x / 10 * 10 + 48);
}
void NL(void)
{
    putchar_unlocked('\n');
}
void SP(void)
{
    putchar_unlocked(' ');
}

void dump_i32(i32 x)
{
    fprintf(stderr, "\033[1;36m%" PRId32 "\033[0m\n", x);
}
void dump_i64(i64 x)
{
    fprintf(stderr, "\033[1;36m%" PRId64 "\033[0m\n", x);
}
void dump_u32(u32 x)
{
    fprintf(stderr, "\033[1;36m%" PRIu32 "\033[0m\n", x);
}
void dump_u64(u64 x)
{
    fprintf(stderr, "\033[1;36m%" PRIu64 "\033[0m\n", x);
}
void dump_i32_array(size_t a_len, i32 *a)
{
    for (size_t i = 0; i < a_len; i++)
    {
        if (i == a_len - 1)
        {
            fprintf(stderr, "\033[1;36m%" PRId32 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRId32 "\033[0m ", a[i]);
        }
    }
}
void dump_i64_array(size_t a_len, i64 *a)
{
    for (size_t i = 0; i < a_len; i++)
    {
        if (i == a_len - 1)
        {
            fprintf(stderr, "\033[1;36m%" PRId64 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRId64 "\033[0m ", a[i]);
        }
    }
}
void dump_u32_array(size_t a_len, u32 *a)
{
    for (size_t i = 0; i < a_len; i++)
    {
        if (i == a_len - 1)
        {
            fprintf(stderr, "\033[1;36m%" PRIu32 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRIu32 "\033[0m ", a[i]);
        }
    }
}
void dump_u64_array(size_t a_len, u64 *a)
{
    for (size_t i = 0; i < a_len; i++)
    {
        if (i == a_len - 1)
        {
            fprintf(stderr, "\033[1;36m%" PRIu64 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRIu64 "\033[0m ", a[i]);
        }
    }
}
void dump_i32_array_range(size_t a_len, i32 *a, size_t l, size_t r)
{
    if (a_len <= r)
    {
        r = a_len - 1;
    }
    if (l > r)
    {
        return;
    }
    for (size_t i = l; i <= r; i++)
    {
        if (i == r)
        {
            fprintf(stderr, "\033[1;36m%" PRId32 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRId32 "\033[0m ", a[i]);
        }
    }
}
void dump_i64_array_range(size_t a_len, i64 *a, size_t l, size_t r)
{
    if (a_len <= r)
    {
        r = a_len - 1;
    }
    if (l > r)
    {
        return;
    }
    for (size_t i = l; i <= r; i++)
    {
        if (i == r)
        {
            fprintf(stderr, "\033[1;36m%" PRId64 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRId64 "\033[0m ", a[i]);
        }
    }
}
void dump_u32_array_range(size_t a_len, u32 *a, size_t l, size_t r)
{
    if (a_len <= r)
    {
        r = a_len - 1;
    }
    if (l > r)
    {
        return;
    }
    for (size_t i = l; i <= r; i++)
    {
        if (i == r)
        {
            fprintf(stderr, "\033[1;36m%" PRIu32 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRIu32 "\033[0m ", a[i]);
        }
    }
}
void dump_u64_array_range(size_t a_len, u64 *a, size_t l, size_t r)
{
    if (a_len <= r)
    {
        r = a_len - 1;
    }
    if (l > r)
    {
        return;
    }
    for (size_t i = l; i <= r; i++)
    {
        if (i == r)
        {
            fprintf(stderr, "\033[1;36m%" PRIu64 "\033[0m\n", a[i]);
        }
        else
        {
            fprintf(stderr, "\033[1;36m%" PRIu64 "\033[0m ", a[i]);
        }
    }
}
void printb_32bit(u32 v)
{
    u32 mask = (u32)1 << (sizeof(v) * CHAR_BIT - 1);
    do
    {
        putchar_unlocked(mask & v ? '1' : '0');
    } while (mask >>= 1);
}
void printb_64bit(u64 v)
{
    u64 mask = (u64)1 << (sizeof(v) * CHAR_BIT - 1);
    do
    {
        putchar_unlocked(mask & v ? '1' : '0');
    } while (mask >>= 1);
}

u32 gcd32(u32 a, u32 b)
{
    if (!a || !b)
        return a | b;
    u32 sh = CTZ32(a | b);
    a >>= CTZ32(a);
    do
    {
        b >>= CTZ32(b);
        if (a > b)
            SWAP(a, b);
        b -= a;
    } while (b);
    return a << sh;
}

u64 gcd64(u64 a, u64 b)
{
    if (!a || !b)
        return a | b;
    u64 sh = CTZ64(a | b);
    a >>= CTZ64(a);
    do
    {
        b >>= CTZ64(b);
        if (a > b)
            SWAP(a, b);
        b -= a;
    } while (b);
    return a << sh;
}

typedef struct
{
    i32 f, s;
    u32 t;
} Bezout32;

typedef struct
{
    i64 f, s;
    u64 t;
} Bezout64;

PRIVATE Bezout32 bezout32(u32 x, u32 y)
{
    bool swap = x < y;
    if (swap)
        SWAP(x, y);
    if (y == 0)
    {
        if (x == 0)
            return (Bezout32){0, 0, 0};
        else if (swap)
            return (Bezout32){0, 1, x};
        else
            return (Bezout32){1, 0, x};
    }
    i32 s0 = 1, s1 = 0, t0 = 0, t1 = 1;
    while (true)
    {
        u32 q = x / y, r = x % y;
        if (r == 0)
        {
            if (swap)
                return (Bezout32){t1, s1, y};
            else
                return (Bezout32){s1, t1, y};
        }
        i32 s2 = s0 - (i32)(q)*s1, t2 = t0 - (i32)(q)*t1;
        x = y, y = r;
        s0 = s1, s1 = s2, t0 = t1, t1 = t2;
    }
}

PRIVATE Bezout64 bezout64(u64 x, u64 y)
{
    bool swap = x < y;
    if (swap)
        SWAP(x, y);
    if (y == 0)
    {
        if (x == 0)
            return (Bezout64){0, 0, 0};
        else if (swap)
            return (Bezout64){0, 1, x};
        else
            return (Bezout64){1, 0, x};
    }
    i64 s0 = 1, s1 = 0, t0 = 0, t1 = 1;
    while (true)
    {
        u64 q = x / y, r = x % y;
        if (r == 0)
        {
            if (swap)
                return (Bezout64){t1, s1, y};
            else
                return (Bezout64){s1, t1, y};
        }
        i64 s2 = s0 - (i64)(q)*s1, t2 = t0 - (i64)(q)*t1;
        x = y, y = r;
        s0 = s1, s1 = s2, t0 = t1, t1 = t2;
    }
}

PUBLIC u32 mod_inverse32(u32 x, u32 mod)
{
    assert(gcd32(x, mod) == 1);
    Bezout32 b = bezout32(x, mod);
    assert(b.t == 1);
    return b.f < 0 ? mod + b.f : (u32)b.f;
}

PUBLIC u64 mod_inverse64(u64 x, u64 mod)
{
    assert(gcd64(x, mod) == 1);
    Bezout64 b = bezout64(x, mod);
    assert(b.t == 1);
    return b.f < 0 ? mod + b.f : (u64)b.f;
}

PRIVATE u32 N_32, N2_32, NI_32, R1_32, R2_32, R3_32;

PUBLIC void Montgomery32(u32 mod)
{
    assert(mod < 1073741824u);
    N_32 = mod;
    N2_32 = mod << 1;
    NI_32 = mod;
    NI_32 *= 2 - NI_32 * mod;
    NI_32 *= 2 - NI_32 * mod;
    NI_32 *= 2 - NI_32 * mod;
    NI_32 *= 2 - NI_32 * mod;
    R1_32 = (u32)(i32)-1 % mod + 1;
    R2_32 = (u64)(i64)-1 % mod + 1;
    R3_32 = (u32)(((u64)R1_32 * (u64)R2_32) % mod);
}
PUBLIC u32 get_mod32(void)
{
    return N_32;
}
PUBLIC u32 get_dmod32(void)
{
    return N2_32;
}
PUBLIC u32 get_one32(void)
{
    return R1_32;
}
PRIVATE u32 mr32(u64 A)
{
    u32 y = (u32)(A >> 32) - (u32)(((u64)((u32)A * NI_32) * N_32) >> 32);
    return (i32)y < 0 ? y + N_32 : y;
}
PUBLIC u32 To32(u32 a)
{
    return mr32((u64)a * R2_32);
}
PUBLIC u32 From32(u32 A)
{
    return mr32((u64)A);
}
PUBLIC u32 Add32(u32 A, u32 B)
{
    A += B - N2_32;
    A += N2_32 & -(A >> 31);
    return A;
}
PUBLIC u32 Sub32(u32 A, u32 B)
{
    A -= B;
    A += N2_32 & -(A >> 31);
    return A;
}
PUBLIC u32 SAdd32(u32 A, u32 B)
{
    A = A >= N_32 ? A % N_32 : A;
    B = B >= N_32 ? B % N_32 : B;
    A += B;
    A -= (A >= N_32 ? N_32 : 0);
    return A;
}
PUBLIC u32 SSub32(u32 A, u32 B)
{
    A = A >= N_32 ? A % N_32 : A;
    B = B >= N_32 ? B % N_32 : B;
    A += (A < B ? N_32 : 0);
    A -= B;
    return A;
}
PUBLIC u32 Min32(u32 A)
{
    return SSub32(0, A);
}
PUBLIC u32 Mul32(u32 A, u32 B)
{
    return mr32((u64)A * B);
}
PUBLIC u32 Square32(u32 A)
{
    return mr32((u64)A * A);
}
PUBLIC u32 Twice32(u32 A)
{
    return (A <<= 1) >= N_32 ? A - N_32 : A;
}
PUBLIC u32 Power32(u32 A, size_t k)
{
    return k ? Mul32(Power32(Square32(A), k >> 1), k & 1 ? A : R1_32) : R1_32;
}
PUBLIC u32 Inverse32(u32 A)
{
    return mr32((u64)R3_32 * mod_inverse32(A, N_32));
}
PUBLIC u32 Div32(u32 A, u32 B)
{
    return Mul32(A, Inverse32(B));
}
PUBLIC u32 Half32(u32 A)
{
    return (A & 1) ? ((A >> 1) + (N_32 >> 1) + 1) : (A >> 1);
}
PUBLIC int Equal32(u32 A, u32 B)
{
    return (((A >= N_32) ? (A - N_32) : A) == ((B >= N_32) ? (B - N_32) : B)) ? 1 : 0;
}
PUBLIC int NotEqual32(u32 A, u32 B)
{
    return (((A >= N_32) ? (A - N_32) : A) != ((B >= N_32) ? (B - N_32) : B)) ? 1 : 0;
}
PUBLIC u32 In32(void)
{
    u32 c = 0;
    u32 a = 0;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        ;
    while (47 < c && c < 58)
    {
        a = a * 10 + c - 48;
        c = getchar_unlocked();
    }
    return To32(a);
}
PUBLIC void Out32(u32 A)
{
    u32 a = From32(A);
    out_u32(a);
}

PRIVATE u64 N_64, N2_64, NI_64, R1_64, R2_64, R3_64;

PUBLIC void Montgomery64(u64 mod)
{
    assert(mod < 4611686018427387904ull);
    N_64 = mod;
    N2_64 = mod << 1;
    NI_64 = mod;
    NI_64 *= 2 - NI_64 * mod;
    NI_64 *= 2 - NI_64 * mod;
    NI_64 *= 2 - NI_64 * mod;
    NI_64 *= 2 - NI_64 * mod;
    NI_64 *= 2 - NI_64 * mod;
    R1_64 = (u64)(i64)-1 % mod + 1;
    R2_64 = (u128)(i128)-1 % mod + 1;
    R3_64 = (u64)(((u128)R1_64 * (u128)R2_64) % mod);
}
PUBLIC u64 get_mod64(void)
{
    return N_64;
}
PUBLIC u64 get_dmod64(void)
{
    return N2_64;
}
PUBLIC u64 get_one64(void)
{
    return R1_64;
}
PRIVATE u64 mr64(u128 A)
{
    u64 y = (u64)(A >> 64) - (u64)(((u128)((u64)A * NI_64) * N_64) >> 64);
    return (i64)y < 0 ? y + N_64 : y;
}
PUBLIC u64 To64(u64 a)
{
    return mr64((u128)a * R2_64);
}
PUBLIC u64 From64(u64 A)
{
    return mr64((u128)A);
}
PUBLIC u64 Add64(u64 A, u64 B)
{
    A += B - N2_64;
    A += N2_64 & -(A >> 63);
    return A;
}
PUBLIC u64 Sub64(u64 A, u64 B)
{
    A -= B;
    A += N2_64 & -(A >> 63);
    return A;
}
PUBLIC u64 SAdd64(u64 A, u64 B)
{
    A = A >= N_64 ? A % N_64 : A;
    B = B >= N_64 ? B % N_64 : B;
    A += B;
    A -= (A >= N_64 ? N_64 : 0);
    return A;
}
PUBLIC u64 SSub64(u64 A, u64 B)
{
    A = A >= N_64 ? A % N_64 : A;
    B = B >= N_64 ? B % N_64 : B;
    A += (A < B ? N_64 : 0);
    A -= B;
    return A;
}
PUBLIC u64 Min64(u64 A)
{
    return SSub64(0, A);
}
PUBLIC u64 Mul64(u64 A, u64 B)
{
    return mr64((u128)A * B);
}
PUBLIC u64 Square64(u64 A)
{
    return mr64((u128)A * A);
}
PUBLIC u64 Twice64(u64 A)
{
    return (A <<= 1) >= N_64 ? A - N_64 : A;
}
PUBLIC u64 Power64(u64 A, size_t k)
{
    return k ? Mul64(Power64(Square64(A), k >> 1), k & 1 ? A : R1_64) : R1_64;
}
PUBLIC u64 Inverse64(u64 A)
{
    return mr64((u128)R3_64 * mod_inverse64(A, N_64));
}
PUBLIC u64 Div64(u64 A, u64 B)
{
    return Mul64(A, Inverse64(B));
}
PUBLIC u64 Half64(u64 A)
{
    return (A & 1) ? ((A >> 1) + (N_64 >> 1) + 1) : (A >> 1);
}
PUBLIC int Equal64(u64 A, u64 B)
{
    return (((A >= N_64) ? (A - N_64) : A) == ((B >= N_64) ? (B - N_64) : B)) ? 1 : 0;
}
PUBLIC int NotEqual64(u64 A, u64 B)
{
    return (((A >= N_64) ? (A - N_64) : A) != ((B >= N_64) ? (B - N_64) : B)) ? 1 : 0;
}
PUBLIC u64 In64()
{
    u64 c = 0;
    u64 a = 0;
    while (c = getchar_unlocked(), c < 48 || c > 57)
        ;
    while (47 < c && c < 58)
    {
        a = a * 10 + c - 48;
        c = getchar_unlocked();
    }
    return To64(a);
}
PUBLIC void Out64(u64 A)
{
    u64 a = From64(A);
    out_u64(a);
}

PRIVATE u64 m_b64, im_b64;
PRIVATE u64 divrem64[2] = {0};

PUBLIC void new_br64(u32 mod)
{
    m_b64 = (u64)mod;
    im_b64 = (~((u64)0ul)) / (u64)mod;
}
PUBLIC u32 get_mod_br32(void)
{
    return (u32)m_b64;
}
PRIVATE void div_rem_br64(u64 lhs)
{
    if (m_b64 == 1)
    {
        divrem64[0] = lhs;
        divrem64[1] = 0;
        return;
    }
    u64 q = (u64)(((u128)lhs * (u128)im_b64) >> 64);
    u64 r = lhs - q * m_b64;
    if (m_b64 <= r)
    {
        r -= m_b64;
        q += 1ul;
    }
    divrem64[0] = q;
    divrem64[1] = r;
}
PUBLIC u32 add_br32(u32 a, u32 b)
{
    a = a >= m_b64 ? a - m_b64 : a;
    b = b >= m_b64 ? b - m_b64 : b;
    a += b;
    a -= (a >= (u32)m_b64 ? (u32)m_b64 : 0);
    return a;
}
PUBLIC u32 sub_br32(u32 a, u32 b)
{
    a = a >= m_b64 ? a - m_b64 : a;
    b = b >= m_b64 ? b - m_b64 : b;
    a += (a < b ? (u32)m_b64 : 0);
    a -= b;
    return a;
}
PUBLIC u32 min_br32(u32 a)
{
    return sub_br32(0, a);
}
PUBLIC u32 mul_br32(u32 a, u32 b)
{
    div_rem_br64((u64)a * b);
    return (u32)divrem64[1];
}
PUBLIC u32 square_br32(u32 a)
{
    div_rem_br64((u64)a * a);
    return (u32)divrem64[1];
}
PUBLIC u32 twice_br32(u32 a)
{
    return mul_br32(a, 2);
}
PUBLIC u32 power_br32(u32 a, size_t k)
{
    return k ? mul_br32(power_br32(square_br32(a), k >> 1), k & 1 ? a : 1) : 1;
}
PUBLIC u32 inverse_br32(u32 a)
{
    if (gcd32(a, m_b64) != 1)
    {
        return 0;
    }
    return mod_inverse32(a, m_b64);
}
PUBLIC u32 div_br32(u32 a, u32 b)
{
    u32 c = inverse_br32(b);
    if (b == 0)
    {
        return (u32)(i32)-1;
    }
    return mul_br32(a, inverse_br32(b));
}
PUBLIC u32 half_br32(u32 a)
{
    return (a & 1) ? ((a >> 1) + (m_b64 >> 1) + 1) : (a >> 1);
}

PRIVATE u128 m_b128, im_b128;
PRIVATE u128 divrem128[2] = {0};

PUBLIC void new_br128(u64 mod)
{
    m_b128 = (u128)mod;
    im_b128 = (~((u128)0ull)) / (u128)mod;
}
PUBLIC u64 get_mod_br64(void)
{
    return (u64)m_b128;
}
PRIVATE void div_rem_br128(u128 lhs)
{
    if (m_b128 == 1)
    {
        divrem128[0] = lhs;
        divrem128[1] = 0;
        return;
    }
    u128 t = (lhs >> 64) * (im_b128 >> 64);
    u128 x = ((lhs & 0xffffffffffffffffull) * (im_b128 & 0xffffffffffffffffull)) >> 64;
    u8 flag;
    u128 auil = (lhs >> 64) * (im_b128 & 0xffffffffffffffffull);
    if (auil <= (u128)((i128)(-1L)) - x)
        flag = 0;
    else
        flag = 1;
    x += auil;
    t += flag;
    u128 aliu = (lhs & 0xffffffffffffffffull) * (im_b128 >> 64);
    if (aliu <= (u128)((i128)(-1L)) - x)
        flag = 0;
    else
        flag = 1;
    x += aliu;
    t += flag;
    u128 q = t + (x >> 64);
    u128 r = lhs - q * m_b128;
    if (m_b128 <= r)
    {
        r -= m_b128;
        q += 1;
    }
    divrem128[0] = q;
    divrem128[1] = r;
}
PUBLIC u64 add_br64(u64 a, u64 b)
{
    a = a >= m_b128 ? a - m_b128 : a;
    b = b >= m_b128 ? b - m_b128 : b;
    a += b;
    a -= (a >= (u64)m_b128 ? (u64)m_b128 : 0);
    return a;
}
PUBLIC u64 sub_br64(u64 a, u64 b)
{
    a = a >= m_b128 ? a - m_b128 : a;
    b = b >= m_b128 ? b - m_b128 : b;
    a += (a < b ? (u64)m_b128 : 0);
    a -= b;
    return a;
}
PUBLIC u64 min_br64(u64 a)
{
    return sub_br64(0, a);
}
PUBLIC u64 mul_br64(u64 a, u64 b)
{
    div_rem_br128((u128)a * b);
    return (u64)divrem128[1];
}
PUBLIC u64 square_br64(u64 a)
{
    div_rem_br128((u128)a * a);
    return (u64)divrem128[1];
}
PUBLIC u64 twice_br64(u64 a)
{
    return mul_br64(a, 2);
}
PUBLIC u64 power_br64(u64 a, size_t k)
{
    return k ? mul_br64(power_br64(square_br64(a), k >> 1), k & 1 ? a : 1) : 1;
}
PUBLIC u64 inverse_br64(u64 a)
{
    if (gcd64(a, m_b128) != 1)
    {
        return 0;
    }
    return mod_inverse64(a, m_b128);
}
PUBLIC u64 div_br64(u64 a, u64 b)
{
    u64 c = inverse_br64(b);
    if (b == 0)
    {
        return (u64)(i64)-1;
    }
    return mul_br64(a, inverse_br64(b));
}
PUBLIC u64 half_br64(u64 a)
{
    return (a & 1) ? ((a >> 1) + (m_b128 >> 1) + 1) : (a >> 1);
}


int isqrt(u64 n)
{
    if (n == 0)
        return 0;

    u64 a = n;
    u64 b = 1;
    u64 c;

    if ((c = (a >> 32)) != 0)
    {
        a = c;
        b <<= 16;
    }
    if ((c = (a >> 16)) != 0)
    {
        a = c;
        b <<= 8;
    }
    if ((c = (a >> 8)) != 0)
    {
        a = c;
        b <<= 4;
    }
    if ((c = (a >> 4)) != 0)
    {
        a = c;
        b <<= 2;
    }
    if ((c = (a >> 2)) != 0)
    {
        a = c;
        b <<= 1;
    }

    if (a <= 1)
        b += b >> 1;
    else
        b <<= 1;

    do
    {
        a = b;
        b = (b + n / b) >> 1;
    } while (b < a);

    return (int)a;
}

enum isqrt_type
{
    Floor = 0,
    Ceil = 1,
    Remain = 2,
};

static inline u64 floor_ceil_remain_isqrt(u64 x, enum isqrt_type mode)
{
    static u64 x_floor_sqrt;
    static u64 x_ceil_sqrt;
    static u64 x_remain_sqrt;
    if (x == 0)
        return 0;
    u32 lz = __builtin_clzll(x);
    u64 n = 32 - (lz >> 1);
    u64 s = (lz >> 1) << 1;
    u64 t = n << 1;
    u128 a = (u128)x;
    u128 b = ((u128)1ull << 62) >> s;
    u128 c = ((u128)1ull << 64) >> s;
    u128 d = (((u128)1ull << 64) - 1) >> s;
    u128 e = ((((u128)1ull << 64) - 1) << 65) >> s;
    for (int _ = 0; _ < n; ++_)
    {
        if (a >= b)
        {
            a -= b;
            b = ((b + b) & e) + c + (b & d);
        }
        else
        {
            b = ((b + b) & e) + (b & d);
        }
        a <<= 2;
    }
    x_floor_sqrt = b >> t;
    x_ceil_sqrt = (a >> t) ? 1ull + (b >> t) : (b >> t);
    x_remain_sqrt = a >> t;

    if (mode == Floor)
    {
        return x_floor_sqrt;
    }
    else if (mode == Ceil)
    {
        return x_ceil_sqrt;
    }
    else
    {
        return x_remain_sqrt;
    }
}

u64 floor_isqrt(u64 n)
{
    return floor_ceil_remain_isqrt(n, Floor);
}

u64 ceil_isqrt(u64 n)
{
    return floor_ceil_remain_isqrt(n, Ceil);
}

u64 remain_isqrt(u64 n)
{
    return floor_ceil_remain_isqrt(n, Remain);
}

bool is_square(u64 n)
{
    if (n <= 1)
        return true;

    if ((0x02030213u >> ((u32)n & 31)) & 1 != 1)
        return false;

    const u64 SQTABLE_MOD4095[64] = {0x2001002010213ul, 0x4200001008028001ul, 0x20000010004ul, 0x80200082010ul, 0x1800008200044029ul, 0x120080000010ul, 0x2200000080410400ul, 0x8100041000200800ul, 0x800004000020100ul, 0x402000400082201ul, 0x9004000040ul, 0x800002000880ul, 0x18002000012000ul, 0x801208ul, 0x26100000804010ul, 0x80000080000002ul, 0x108040040101045ul, 0x20c00004000102ul, 0x400000100c0010ul, 0x1300000040208ul, 0x804000020010000ul, 0x1008402002400080ul, 0x201001000200040ul, 0x4402000000806000ul, 0x10402000000ul, 0x1040008001200801ul, 0x4080000000020400ul, 0x10083080000002ul, 0x8220140000040000ul, 0x800084020100000ul, 0x80010400010000ul, 0x1200020108008060ul, 0x180000000ul, 0x400002400000018ul, 0x4241000200ul, 0x100800000000ul, 0x10201008400483ul, 0xc008000208201000ul, 0x800420000100ul, 0x2010002000410ul, 0x28041000000ul, 0x4010080000024ul, 0x400480010010080ul, 0x200040028000008ul, 0x100810084020ul, 0x20c0401000080000ul, 0x1000240000220000ul, 0x4000020800ul, 0x410000000480000ul, 0x8004008000804201ul, 0x806020000104000ul, 0x2080002000211000ul, 0x1001008001000ul, 0x20000010024000ul, 0x480200002040000ul, 0x48200044008000ul, 0x100000000010080ul, 0x80090400042ul, 0x41040200800200ul, 0x4000020100110ul, 0x2000400082200010ul, 0x1008200000000040ul, 0x2004800002ul, 0x2002010000080ul};
    size_t p = n % 4095;
    if ((SQTABLE_MOD4095[p >> 6] >> (p & 63)) & 1 != 1)
        return false;

    u64 newton_sqrt;
    size_t k = 32 - (CLZ64(n - 1) >> 1);
    u64 s = (u64)(1ul) << k;
    u64 t = (s + (n >> k)) >> 1;
    while (t < s)
    {
        s = t;
        t = (s + n / s) >> 1;
    }
    newton_sqrt = s;
    if (newton_sqrt * newton_sqrt != n)
        return false;

    return true;
}


int jacobi_symbol32(i32 a, i32 n)
{
    int j = 1;
    while (a)
    {
        if (a < 0)
        {
            a = -a;
            if ((n & 3) == 3)
                j = -j;
        }
        int s = CTZ32(a);
        a >>= s;
        if (((n & 7) == 3 || (n & 7) == 5) && (s & 1))
            j = -j;
        if ((a & n & 3) == 3)
            j = -j;
        SWAP(a, n);
        a %= n;
        if (a > n / 2)
            a -= n;
    }
    return n == 1 ? j : 0;
}

int jacobi_symbol64(i64 a, i64 n)
{
    int j = 1;
    while (a)
    {
        if (a < 0)
        {
            a = -a;
            if ((n & 3) == 3)
                j = -j;
        }
        int s = CTZ64(a);
        a >>= s;
        if (((n & 7) == 3 || (n & 7) == 5) && (s & 1))
            j = -j;
        if ((a & n & 3) == 3)
            j = -j;
        SWAP(a, n);
        a %= n;
        if (a > n / 2)
            a -= n;
    }
    return n == 1 ? j : 0;
}


PRIVATE bool _miller_rabin(u64 n, size_t bases_len, u64 bases[])
{
    u64 s = CTZ64(n - 1);
    u64 d = (n - 1) >> s;
    Montgomery64(n);
    for (size_t i = 0; i < bases_len; ++i)
    {
        if (n <= bases[i])
            return true;
        u64 a = Power64(To64(bases[i]), d);
        if (a == get_one64())
            continue;
        u64 r = 1;
        while (a != n - get_one64())
        {
            if (r == s)
                return false;
            a = Square64(a);
            ++r;
        }
    }
    return true;
}

PRIVATE bool _miller_rabin_br(u64 n)
{
    new_br128(n);
    u64 s = CTZ64(n - 1);
    u64 d = (n - 1) >> s;
    u64 bases[7] = {2ul, 325ul, 9375ul, 28178ul, 450775ul, 9780504ul, 1795265022ul};
    for (size_t i = 0; i < 7; ++i)
    {
        u64 a = power_br64(bases[i], d);
        if (a == 1)
            continue;
        u64 r = 1;
        while (a != n - 1)
        {
            if (r == s)
                return false;
            a = square_br64(a);
            ++r;
        }
    }
    return true;
}

PUBLIC bool miller_rabin(u64 n)
{
    if (n < 64ull)
        return (1ull << n) & 2891462833508853932ull;

    if (!(n & 1))
        return false;

    if (n < 1073741824ull)
    {
        u64 bases[3] = {2ul, 7ul, 61ul};
        return _miller_rabin(n, 3, bases);
    }

    if (n < 4611686018427387904ull)
    {
        u64 bases[7] = {2ul, 325ul, 9375ul, 28178ul, 450775ul, 9780504ul, 1795265022ul};
        return _miller_rabin(n, 7, bases);
    }

    return _miller_rabin_br(n);
}

PUBLIC bool baillie_psw(u64 n)
{
    if (n < 64ull)
        return (1ull << n) & 2891462833508853932ull;

    if (!(n & 1))
        return false;

    new_br128(n);

    {
        u64 d = (n - 1) << CLZ64(n - 1);
        u64 t = 2ull;
        for (d <<= 1; d; d <<= 1)
        {
            t = square_br64(t);
            if (d >> 63)
                t = twice_br64(t);
        }
        if (t != 1)
        {
            u64 x = LSBit(n - 1);
            u64 rev = n - 1;
            for (x >>= 1; t != rev; x >>= 1)
            {
                if (x == 0)
                    return false;
                t = square_br64(t);
            }
        }
    }

    {
        i64 D = 5;
        for (int i = 0; jacobi_symbol64(D, n) != -1 && i < 64; ++i)
        {
            if (i == 32 && is_square(n))
            {
                return false;
            }
            if (i & 1)
                D -= 2;
            else
                D += 2;
            D = -D;
        }
        u64 Q = (D < 0) ? ((1 - D) / 4 % n) : (n - (D - 1) / 4 % n);
        u64 u = 1, v = 1, Qn = Q;
        u64 k = (n + 1) << CLZ64(n + 1);
        D %= (i64)n;
        D = (D < 0) ? (D + n) : D;
        for (k <<= 1; k; k <<= 1)
        {
            u = mul_br64(u, v);
            v = sub_br64(square_br64(v), twice_br64(Qn));
            Qn = square_br64(Qn);
            if (k >> 63)
            {
                u64 uu = add_br64(u, v);
                uu = half_br64(uu);
                v = half_br64(add_br64(mul_br64(D, u), v));
                u = uu;
                Qn = mul_br64(Qn, Q);
            }
        }
        if (u == 0 || v == 0)
            return true;
        u64 x = (n + 1) & ~n;
        for (x >>= 1; x; x >>= 1)
        {
            u = mul_br64(u, v);
            v = sub_br64(square_br64(v), twice_br64(Qn));
            if (v == 0)
                return true;
            Qn = square_br64(Qn);
        }
    }
    return false;
}




int main(int argc, char *argv[])
{
    int Q = in_i32();
    while (Q--)
    {
        u64 x = in_u64();
        out_u64(x);
        SP();
        putchar_unlocked(miller_rabin(x) ? '1' : '0');
        NL();
    }
    return 0;
}
yukicoder

結果

テストケース

ソースコード
