qzh/se-algo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

在method里面统一定义缓存变量，所有函数全塞到一起，大小驼峰下划线全大写变量名混用

Browse Source
This commit is contained in:
zcy
2025-09-09 01:30:35 +08:00
parent fe4e0ff4e6
commit dec1fecc16
7 changed files with 3397 additions and 1371 deletions
Download Patch File Download Diff File Expand all files Collapse all files

978
Project/Src/com/cscn/Zuc256Util.java
View File

@@ -1,497 +1,497 @@
package com.cscn;
import javacard.framework.Util;
import javacard.framework.JCSystem;
/**
* 辅助工具：装载/存储、位运算、线性变换、打印等。
*/
public final class Zuc256Util {
private Zuc256Util() {}
// /** 辅助方法：将字节数组转换为32位整数 */
// public static int getU32(byte[] p, int offset) {
// return ((p[offset] & 0xFF) << 24) |
// ((p[offset + 1] & 0xFF) << 16) |
// ((p[offset + 2] & 0xFF) << 8) |
// (p[offset + 3] & 0xFF);
//package com.cscn;
//
//import javacard.framework.Util;
//import javacard.framework.JCSystem;
//
///**
// * 辅助工具：装载/存储、位运算、线性变换、打印等。
// */
//public final class Zuc256Util {
//
// private Zuc256Util() {}
//
//// /** 辅助方法：将字节数组转换为32位整数 */
//// public static int getU32(byte[] p, int offset) {
//// return ((p[offset] & 0xFF) << 24) |
//// ((p[offset + 1] & 0xFF) << 16) |
//// ((p[offset + 2] & 0xFF) << 8) |
//// (p[offset + 3] & 0xFF);
//// }
// /** 辅助方法：从字节数组取出 32 位整数，存放到 short[2] (lo, hi) */
// public static void getU32(byte[] p, short offset, short[] out32 /* len=2 */) {
// out32[0] = (short) (((p[(short)(offset + 2)] & 0xFF) << 8) | (p[(short)(offset + 3)] & 0xFF)); //低16位
// out32[1] = (short) (((p[offset] & 0xFF) << 8) | (p[(short)(offset + 1)] & 0xFF)); //高16位
// }
/** 辅助方法：从字节数组取出 32 位整数，存放到 short[2] (lo, hi) */
public static void getU32(byte[] p, short offset, short[] out32 /* len=2 */) {
out32[0] = (short) (((p[(short)(offset + 2)] & 0xFF) << 8) | (p[(short)(offset + 3)] & 0xFF)); //低16位
out32[1] = (short) (((p[offset] & 0xFF) << 8) | (p[(short)(offset + 1)] & 0xFF)); //高16位
}
// /** 辅助方法：将32位整数转换为字节数组 */
// public static void putU32(byte[] p, int offset, int v) {
// p[offset] = (byte) (v >> 24);
// p[offset + 1] = (byte) (v >> 16);
// p[offset + 2] = (byte) (v >> 8);
// p[offset + 3] = (byte) v;
//
//
//// /** 辅助方法：将32位整数转换为字节数组 */
//// public static void putU32(byte[] p, int offset, int v) {
//// p[offset] = (byte) (v >> 24);
//// p[offset + 1] = (byte) (v >> 16);
//// p[offset + 2] = (byte) (v >> 8);
//// p[offset + 3] = (byte) v;
//// }
// /** 辅助方法：将32位整数(vlo=低16位, vhi=高16位)写入字节数组 */
// public static void putU32(byte[] p, short offset, short vlo, short vhi) {
// // 写高16位
// p[offset] = (byte) ((vhi >> 8) & 0xFF);
// p[(short)(offset + 1)] = (byte) (vhi & 0xFF);
//
// // 写低16位
// p[(short)(offset + 2)] = (byte) ((vlo >> 8) & 0xFF);
// p[(short)(offset + 3)] = (byte) (vlo & 0xFF);
// }
/** 辅助方法：将32位整数(vlo=低16位, vhi=高16位)写入字节数组 */
public static void putU32(byte[] p, short offset, short vlo, short vhi) {
// 写高16位
p[offset] = (byte) ((vhi >> 8) & 0xFF);
p[(short)(offset + 1)] = (byte) (vhi & 0xFF);
// 写低16位
p[(short)(offset + 2)] = (byte) ((vlo >> 8) & 0xFF);
p[(short)(offset + 3)] = (byte) (vlo & 0xFF);
}
// === 31/32 位运算 ===
// /** 31位加法 */
// public static int add31(int a, int b) {
// long sum = (long)a + b;
// return (int) ((sum & 0x7FFFFFFF) + (sum >> 31));
// }
/** 31位加法: (a+b) mod (2^31 - 1)
* 输入: a_lo=低16位, a_hi=高15位
* b_lo=低16位, b_hi=高15位
* 输出: out[0]=lo, out[1]=hi
*/
public static void add31(short a_lo, short a_hi, short b_lo, short b_hi, short[] out /* len==2 */) {
// ---- 低16位相加 ----
short lo = (short)(a_lo + b_lo);
short carry = (short)(
( ( (short)( (a_lo & b_lo) | ((a_lo | b_lo) & (short)~lo) ) ) & (short)0x8000 ) != 0
? 1 : 0
);
// ---- 高15位相加 + 进位 ----
short hi_raw = (short)((short)((a_hi & 0x7FFF) + (b_hi & 0x7FFF)) + carry);
// 提取第31位（hi_raw bit15）
short topbit = (short)((hi_raw >>> 15) & 1);
short hi = (short)(hi_raw & 0x7FFF); // 保留15位
// ---- 若第31位=1，再+1 ----
if (topbit == 1) {
short lo2 = (short)(lo + 1);
short c2 = (short)((lo2 == 0) ? 1 : 0); // lo溢出时进位
lo = lo2;
hi = (short)((hi + c2) & 0x7FFF);
}
out[0] = lo;
out[1] = hi;
}
// /** 31位旋转 */
// public static int rot31(int a, int k) {
// return ((a << k) | (a >>> (31 - k))) & 0x7FFFFFFF;
// }
/** 31位循环左移: (a <<< k) mod (2^31 -1)
* 输入: a_lo=低16位, a_hi=高15位
* 输出: out[0]=lo, out[1]=hi
*/
public static void rot31(short a_lo, short a_hi, short k, short[] out /* len==2 */) {
k = (short)(k % 31); // 限制在 0..30
if (k == 0) {
out[0] = a_lo;
out[1] = (short)(a_hi & 0x7FFF);
return;
}
// 拆成 31 位数组 [bit0..bit30]
short[] bits = JCSystem.makeTransientShortArray((short)31, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
for (short i = 0; i < 16; i++) {
bits[i] = (short)((a_lo >>> i) & 1);
}
for (short i = 0; i < 15; i++) {
bits[(short)(16 + i)] = (short)((a_hi >>> i) & 1);
}
// 旋转
short[] resBits = JCSystem.makeTransientShortArray((short)31, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
for (short i = 0; i < 31; i++) {
short j = (short)((i + k) % 31);
resBits[j] = bits[i];
}
// 拼回 lo, hi
short lo = 0;
for (short i = 0; i < 16; i++) {
lo = (short)(lo | (resBits[i] << i));
}
short hi = 0;
for (short i = 0; i < 15; i++) {
hi = (short)(hi | (resBits[(short)(16 + i)] << i));
}
out[0] = lo;
out[1] = hi;
}
// /** 32位旋转 */
// public static int rot32(int a, int k) {
// return (a << k) | (a >>> (32 - k));
// }
/** 32位循环左移: (a<<<k) */
public static void rot32(short a_lo, short a_hi, short k, short[] out /*len==2*/) {
k = (short)(k & 31); // 0..31
short lo = a_lo, hi = a_hi, nw_hi, nw_lo;
while (k > 0) {
// 先做 1 位循环左移
// 注意：short 在 >>> 时会先提升为 int，所以下面都再用 &1 取最低位，避免符号扩展影响
nw_hi = (short)((hi << 1) | ((lo >>> 15) & 1));
nw_lo = (short)((lo << 1) | ((hi >>> 15) & 1));
hi = nw_hi;
lo = nw_lo;
k--;
}
out[0] = lo; // 低16位
out[1] = hi; // 高16位
}
// /**
// * L1函数
//
//
// // === 31/32 位运算 ===
//
//// /** 31位加法 */
//// public static int add31(int a, int b) {
//// long sum = (long)a + b;
//// return (int) ((sum & 0x7FFFFFFF) + (sum >> 31));
//// }
// /** 31位加法: (a+b) mod (2^31 - 1)
// * 输入: a_lo=低16位, a_hi=高15位
// * b_lo=低16位, b_hi=高15位
// * 输出: out[0]=lo, out[1]=hi
// */
// public static int L1(int x) {
// return x ^ rot32(x, 2) ^ rot32(x, 10) ^ rot32(x, 18) ^ rot32(x, 24);
// public static void add31(short a_lo, short a_hi, short b_lo, short b_hi, short[] out /* len==2 */) {
// // ---- 低16位相加 ----
// short lo = (short)(a_lo + b_lo);
// short carry = (short)(
// ( ( (short)( (a_lo & b_lo) | ((a_lo | b_lo) & (short)~lo) ) ) & (short)0x8000 ) != 0
// ? 1 : 0
// );
// // ---- 高15位相加 + 进位 ----
// short hi_raw = (short)((short)((a_hi & 0x7FFF) + (b_hi & 0x7FFF)) + carry);
//
// // 提取第31位（hi_raw bit15）
// short topbit = (short)((hi_raw >>> 15) & 1);
// short hi = (short)(hi_raw & 0x7FFF); // 保留15位
//
// // ---- 若第31位=1，再+1 ----
// if (topbit == 1) {
// short lo2 = (short)(lo + 1);
// short c2 = (short)((lo2 == 0) ? 1 : 0); // lo溢出时进位
// lo = lo2;
// hi = (short)((hi + c2) & 0x7FFF);
// }
//
// out[0] = lo;
// out[1] = hi;
// }
/**
* L1函数: x ^ (x<<<2) ^ (x<<<10) ^ (x<<<18) ^ (x<<<24)
* 输入: x_lo, x_hi
* 输出: out[0]=lo, out[1]=hi
*/
public static void L1(short x_lo, short x_hi, short[] out /*len==2*/) {
short[] t = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
short[] acc = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// acc = x
acc[0] = x_lo;
acc[1] = x_hi;
// acc ^= rot32(x, 2)
rot32(x_lo, x_hi, (short)2, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 10)
rot32(x_lo, x_hi, (short)10, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 18)
rot32(x_lo, x_hi, (short)18, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 24)
rot32(x_lo, x_hi, (short)24, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
out[0] = acc[0];
out[1] = acc[1];
}
// /**
// * L2函数
//
//
//// /** 31位旋转 */
//// public static int rot31(int a, int k) {
//// return ((a << k) | (a >>> (31 - k))) & 0x7FFFFFFF;
//// }
// /** 31位循环左移: (a <<< k) mod (2^31 -1)
// * 输入: a_lo=低16位, a_hi=高15位
// * 输出: out[0]=lo, out[1]=hi
// */
// public static int L2(int x) {
// return x ^ rot32(x, 8) ^ rot32(x, 14) ^ rot32(x, 22) ^ rot32(x, 30);
// public static void rot31(short a_lo, short a_hi, short k, short[] out /* len==2 */) {
// k = (short)(k % 31); // 限制在 0..30
// if (k == 0) {
// out[0] = a_lo;
// out[1] = (short)(a_hi & 0x7FFF);
// return;
// }
//
// // 拆成 31 位数组 [bit0..bit30]
// short[] bits = JCSystem.makeTransientShortArray((short)31, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// for (short i = 0; i < 16; i++) {
// bits[i] = (short)((a_lo >>> i) & 1);
// }
// for (short i = 0; i < 15; i++) {
// bits[(short)(16 + i)] = (short)((a_hi >>> i) & 1);
// }
//
// // 旋转
// short[] resBits = JCSystem.makeTransientShortArray((short)31, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// for (short i = 0; i < 31; i++) {
// short j = (short)((i + k) % 31);
// resBits[j] = bits[i];
// }
//
// // 拼回 lo, hi
// short lo = 0;
// for (short i = 0; i < 16; i++) {
// lo = (short)(lo | (resBits[i] << i));
// }
// short hi = 0;
// for (short i = 0; i < 15; i++) {
// hi = (short)(hi | (resBits[(short)(16 + i)] << i));
// }
//
// out[0] = lo;
// out[1] = hi;
// }
/**
* L2函数: x ^ (x<<<8) ^ (x<<<14) ^ (x<<<22) ^ (x<<<30)
* 输入: x_lo, x_hi
* 输出: out[0]=lo, out[1]=hi
*/
public static void L2(short x_lo, short x_hi, short[] out /*len==2*/) {
short[] t = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
short[] acc = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// acc = x
acc[0] = x_lo;
acc[1] = x_hi;
// acc ^= rot32(x, 8)
rot32(x_lo, x_hi, (short)8, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 14)
rot32(x_lo, x_hi, (short)14, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 22)
rot32(x_lo, x_hi, (short)22, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
// acc ^= rot32(x, 30)
rot32(x_lo, x_hi, (short)30, t);
acc[0] ^= t[0];
acc[1] ^= t[1];
out[0] = acc[0];
out[1] = acc[1];
}
// /** 创建31位无符号整数 */
// public static int makeU31(int a, int b, int c, int d) {
// return (((a & 0xFF) << 23) |
// ((b & 0xFF) << 16) |
// ((c & 0xFF) << 8) |
// (d & 0xFF)) & 0x7FFFFFFF;
//
//
//// /** 32位旋转 */
//// public static int rot32(int a, int k) {
//// return (a << k) | (a >>> (32 - k));
//// }
// /** 32位循环左移: (a<<<k) */
// public static void rot32(short a_lo, short a_hi, short k, short[] out /*len==2*/) {
// k = (short)(k & 31); // 0..31
// short lo = a_lo, hi = a_hi, nw_hi, nw_lo;
// while (k > 0) {
// // 先做 1 位循环左移
// // 注意：short 在 >>> 时会先提升为 int，所以下面都再用 &1 取最低位，避免符号扩展影响
// nw_hi = (short)((hi << 1) | ((lo >>> 15) & 1));
// nw_lo = (short)((lo << 1) | ((hi >>> 15) & 1));
// hi = nw_hi;
// lo = nw_lo;
// k--;
// }
// out[0] = lo; // 低16位
// out[1] = hi; // 高16位
// }
/** 创建31位无符号整数，结果放到 out[0]=lo, out[1]=hi(15位) */
public static void makeU31(short a, short b, short c, short d, short[] out /*len==2*/) {
// 四个字节
short b0 = (short)(a & 0xFF); // 最高字节
short b1 = (short)(b & 0xFF);
short b2 = (short)(c & 0xFF);
short b3 = (short)(d & 0xFF); // 最低字节
// 拼成 32 位: b0<<24 | b1<<16 | b2<<8 | b3
// lo = 低16位
out[0] = (short)((b2 << 8) | b3);
// hi = 高15位（丢弃 bit31）
out[1] = (short)((b0 << 7) | b1);
}
// /** 创建32位无符号整数 */
// public static int makeU32(int a, int b, int c, int d) {
// return ((a & 0xFF) << 24) |
// ((b & 0xFF) << 16) |
// ((c & 0xFF) << 8) |
// (d & 0xFF);
//
//
//// /**
//// * L1函数
//// */
//// public static int L1(int x) {
//// return x ^ rot32(x, 2) ^ rot32(x, 10) ^ rot32(x, 18) ^ rot32(x, 24);
//// }
// /**
// * L1函数: x ^ (x<<<2) ^ (x<<<10) ^ (x<<<18) ^ (x<<<24)
// * 输入: x_lo, x_hi
// * 输出: out[0]=lo, out[1]=hi
// */
// public static void L1(short x_lo, short x_hi, short[] out /*len==2*/) {
// short[] t = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// short[] acc = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
//
// // acc = x
// acc[0] = x_lo;
// acc[1] = x_hi;
//
// // acc ^= rot32(x, 2)
// rot32(x_lo, x_hi, (short)2, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 10)
// rot32(x_lo, x_hi, (short)10, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 18)
// rot32(x_lo, x_hi, (short)18, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 24)
// rot32(x_lo, x_hi, (short)24, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// out[0] = acc[0];
// out[1] = acc[1];
// }
/** 创建32位无符号整数，结果放到 out[0]=lo, out[1]=hi */
public static void makeU32(short a, short b, short c, short d, short[] out /*len==2*/) {
// 四个字节
short b0 = (short)(a & 0xFF); // 最高字节
short b1 = (short)(b & 0xFF);
short b2 = (short)(c & 0xFF);
short b3 = (short)(d & 0xFF); // 最低字节
// lo = 低16位
out[0] = (short)((b2 << 8) | b3);
// hi = 高16位
out[1] = (short)((b0 << 8) | b1);
}
/** 提取IV */
public static void extractIv(byte[] input25Byte, byte[] output23Byte) {
if (input25Byte == null || output23Byte == null) return;
// 复制前17字节
Util.arrayCopyNonAtomic(input25Byte, (short)0, output23Byte, (short)0, (short)17);
// 处理剩余8字节
byte[] src = JCSystem.makeTransientByteArray((short)8, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
for (short i = 0; i < 8; i++) {
src[i] = (byte) (input25Byte[(short)(17 + i)] & 0x3F);
}
output23Byte[17] = (byte) ((src[0] << 2) | (src[1] >>> 4));
output23Byte[18] = (byte) (((src[1] & 0x0F) << 4) | (src[2] >>> 2));
output23Byte[19] = (byte) (((src[2] & 0x03) << 6) | src[3]);
output23Byte[20] = (byte) ((src[4] << 2) | (src[5] >>> 4));
output23Byte[21] = (byte) (((src[5] & 0x0F) << 4) | (src[6] >>> 2));
output23Byte[22] = (byte) (((src[6] & 0x03) << 6) | src[7]);
}
/**
* 32位加法: (a_hi:a_lo) + (b_hi:b_lo)
* out[0] = lo, out[1] = hi
*/
static void add32(short a_lo, short a_hi,
short b_lo, short b_hi,
short[] out /*len=2*/) {
// ---- 低16位 ----
short lo_low = (short)((a_lo & 0x00FF) + (b_lo & 0x00FF));
short carry0 = (short)(((a_lo & 0x00FF) + (b_lo & 0x00FF)) >>> 8);
short a_lo_hi = (short)((a_lo >>> 8) & 0x00FF);
short b_lo_hi = (short)((b_lo >>> 8) & 0x00FF);
short lo_high = (short)(a_lo_hi + b_lo_hi + carry0);
short carry1 = (short)(lo_high >>> 8);
short lo_res = (short)((lo_high << 8) | (lo_low & 0x00FF));
// ---- 高16位 ----
short hi_low = (short)((a_hi & 0x00FF) + (b_hi & 0x00FF) + carry1);
short carry2 = (short)(hi_low >>> 8);
short a_hi_hi = (short)((a_hi >>> 8) & 0x00FF);
short b_hi_hi = (short)((b_hi >>> 8) & 0x00FF);
short hi_high = (short)(a_hi_hi + b_hi_hi + carry2);
short hi_res = (short)((hi_high << 8) | (hi_low & 0x00FF));
// ---- 输出 ----
out[0] = lo_res;
out[1] = hi_res;
}
/**
* 32位加法 + 返回进位（只用 short）
* 输入: (a_hi:a_lo) + (b_hi:b_lo)
* 输出: out[0]=lo, out[1]=hi
* 返回: 最终进位(0/1)
*/
static short add32_with_carry(short a_lo, short a_hi,
short b_lo, short b_hi,
short[] out /* len=2 */) {
// ---- 低16位：分两段8位相加 ----
short s0 = (short)((a_lo & (short)0x00FF) + (b_lo & (short)0x00FF)); // 0..510
short c0 = (short)(s0 >>> 8); // 0/1
short s1 = (short)(((a_lo >>> 8) & (short)0x00FF)
+ ((b_lo >>> 8) & (short)0x00FF)
+ c0); // 0..511
short c1 = (short)(s1 >>> 8); // 0/1
short lo = (short)((s1 << 8) | (s0 & (short)0x00FF));
// ---- 高16位：再分两段8位相加，并加上 c1 ----
short s2 = (short)((a_hi & (short)0x00FF) + (b_hi & (short)0x00FF) + c1);
short c2 = (short)(s2 >>> 8); // 0/1
short s3 = (short)(((a_hi >>> 8) & (short)0x00FF)
+ ((b_hi >>> 8) & (short)0x00FF)
+ c2); // 0..511
short c3 = (short)(s3 >>> 8); // 最终进位 0/1
short hi = (short)((s3 << 8) | (s2 & (short)0x00FF));
out[0] = lo;
out[1] = hi;
return (short)(c3 & 1);
}
/**
* 64位加法: a4 + b4 -> a4
* 输入输出: short[4]，低到高 (a[0]=lo16, a[1]=hi16, a[2]=lo16 of high dword, a[3]=hi16 of high dword)
*/
static void add64(short[] a, short[] b) {
short[] tmp = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// 低 32 位
short carry = add32_with_carry(a[0], a[1], b[0], b[1], tmp);
a[0] = tmp[0];
a[1] = tmp[1];
// 高 32 位 + carry
add32((short)(a[2] + (short)(carry & (short)0x0001)), a[3], b[2], b[3], tmp);
a[2] = tmp[0];
a[3] = tmp[1];
}
// 32位异或
public static void xor32(short a_lo, short a_hi, short b_lo, short b_hi, short[] out /*len==2*/) {
out[0] = (short)(a_lo ^ b_lo);
out[1] = (short)(a_hi ^ b_hi);
}
/**
* 把32位数 b (b[0]=lo, b[1]=hi) 左移 k 位 (0 <= k < 32)，
* 结果放到64位数 a (a[0]=最低16位 ... a[3]=最高16位)。
*/
static void create_64b_from_32b(short[] a/*len=4*/, short[] b/*len=2*/, short k) {
short a0 = b[0], a1 = b[1], a2 = 0, a3 = 0;
if (k >= 16) {
a3 = a2; // 0
a2 = a1; // 原 hi16
a1 = a0; // 原 lo16
a0 = 0;
k = (short)(k - 16);
}
while (k > 0) {
short c0 = (short)((a0 >>> 15) & 1);
short c1 = (short)((a1 >>> 15) & 1);
short c2 = (short)((a2 >>> 15) & 1);
a3 = (short)((a3 << 1) | c2);
a2 = (short)((a2 << 1) | c1);
a1 = (short)((a1 << 1) | c0);
a0 = (short)(a0 << 1);
k--;
}
a[0] = a0; a[1] = a1; a[2] = a2; a[3] = a3;
}
/**
* (A & 0x7FFFFFFF)，结果放在 out[4]，只保留低32位并清掉最高bit。
*/
static void and64_7FFFFFFF_to32(short[] A, short[] out) {
out[0] = A[0]; // lo16
out[1] = (short)(A[1] & 0x7FFF); // hi16 (清除最高bit)
out[2] = 0;
out[3] = 0;
}
/**
* 64位无符号右移 31 位
* 输入: A[0..3] (short[4], A[0]最低16位)
* 输出: out[0..3]
*/
static void shr64u_31(short[] A, short[] out) {
// 先拼出 64bit 的逻辑，逐段右移
// A >>> 31 = (A >>> 16) >>> 15
// 先右移 16，相当于丢掉 A[0]，整体右移一半字
out[0] = A[1]; // 原 A[1] -> 新低16位
out[1] = A[2]; // 原 A[2]
out[2] = A[3]; // 原 A[3]
out[3] = 0; // 高位补0
// 再右移 15 位
short c0 = (short)((out[0] & (short)0xFFFF) >>> 15); // out[0] 最后一位变进位
short c1 = (short)((out[1] & (short)0xFFFF) >>> 15);
short c2 = (short)((out[2] & (short)0xFFFF) >>> 15);
out[0] = (short)((c0 & 0x0001) | (out[1] << 1));
out[1] = (short)((c1 & 0x0001) | (out[2] << 1));
out[2] = (short)(c2 & 0x0001);
}
/**
* 32位无符号右移 1 位
* 输入: lo,hi (short) 表示 32 位数 (hi:高16位, lo:低16位)
* 输出: out[0]=lo, out[1]=hi
*/
static void shr32u1(short lo, short hi, short[] out) {
// >>>1：先处理低16位
short nwLo = (short)(((((lo & (short)0xFFFF) >>> 1) & (short)0x7FFF)) | ((hi & 0x0001) << 15));
short nwHi = (short)(((hi & (short)0xFFFF) >>> 1) & (short)0x7FFF);
out[0] = nwLo;
out[1] = nwHi;
}
/** 打印/*十六进制（调试用，TODO 生产/JC 环境可移除） *//*
public static void printHex(String label, byte[] data, int len) {
System.out.print(label + ": ");
for (int i = 0; i < len; i++) {
System.out.printf("%02x ", data[i] & 0xFF);
}
System.out.println();
}*/
}
//
//
//// /**
//// * L2函数
//// */
//// public static int L2(int x) {
//// return x ^ rot32(x, 8) ^ rot32(x, 14) ^ rot32(x, 22) ^ rot32(x, 30);
//// }
// /**
// * L2函数: x ^ (x<<<8) ^ (x<<<14) ^ (x<<<22) ^ (x<<<30)
// * 输入: x_lo, x_hi
// * 输出: out[0]=lo, out[1]=hi
// */
// public static void L2(short x_lo, short x_hi, short[] out /*len==2*/) {
// short[] t = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// short[] acc = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
//
// // acc = x
// acc[0] = x_lo;
// acc[1] = x_hi;
//
// // acc ^= rot32(x, 8)
// rot32(x_lo, x_hi, (short)8, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 14)
// rot32(x_lo, x_hi, (short)14, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 22)
// rot32(x_lo, x_hi, (short)22, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// // acc ^= rot32(x, 30)
// rot32(x_lo, x_hi, (short)30, t);
// acc[0] ^= t[0];
// acc[1] ^= t[1];
//
// out[0] = acc[0];
// out[1] = acc[1];
// }
//
//
//// /** 创建31位无符号整数 */
//// public static int makeU31(int a, int b, int c, int d) {
//// return (((a & 0xFF) << 23) |
//// ((b & 0xFF) << 16) |
//// ((c & 0xFF) << 8) |
//// (d & 0xFF)) & 0x7FFFFFFF;
//// }
// /** 创建31位无符号整数，结果放到 out[0]=lo, out[1]=hi(15位) */
// public static void makeU31(short a, short b, short c, short d, short[] out /*len==2*/) {
// // 四个字节
// short b0 = (short)(a & 0xFF); // 最高字节
// short b1 = (short)(b & 0xFF);
// short b2 = (short)(c & 0xFF);
// short b3 = (short)(d & 0xFF); // 最低字节
//
// // 拼成 32 位: b0<<24 | b1<<16 | b2<<8 | b3
// // lo = 低16位
// out[0] = (short)((b2 << 8) | b3);
//
// // hi = 高15位（丢弃 bit31）
// out[1] = (short)((b0 << 7) | b1);
// }
//
//
//// /** 创建32位无符号整数 */
//// public static int makeU32(int a, int b, int c, int d) {
//// return ((a & 0xFF) << 24) |
//// ((b & 0xFF) << 16) |
//// ((c & 0xFF) << 8) |
//// (d & 0xFF);
//// }
// /** 创建32位无符号整数，结果放到 out[0]=lo, out[1]=hi */
// public static void makeU32(short a, short b, short c, short d, short[] out /*len==2*/) {
// // 四个字节
// short b0 = (short)(a & 0xFF); // 最高字节
// short b1 = (short)(b & 0xFF);
// short b2 = (short)(c & 0xFF);
// short b3 = (short)(d & 0xFF); // 最低字节
//
// // lo = 低16位
// out[0] = (short)((b2 << 8) | b3);
//
// // hi = 高16位
// out[1] = (short)((b0 << 8) | b1);
// }
//
//
//
// /** 提取IV */
// public static void extractIv(byte[] input25Byte, byte[] output23Byte) {
// if (input25Byte == null || output23Byte == null) return;
//
// // 复制前17字节
// Util.arrayCopyNonAtomic(input25Byte, (short)0, output23Byte, (short)0, (short)17);
//
//
// // 处理剩余8字节
// byte[] src = JCSystem.makeTransientByteArray((short)8, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
// for (short i = 0; i < 8; i++) {
// src[i] = (byte) (input25Byte[(short)(17 + i)] & 0x3F);
// }
//
// output23Byte[17] = (byte) ((src[0] << 2) | (src[1] >>> 4));
// output23Byte[18] = (byte) (((src[1] & 0x0F) << 4) | (src[2] >>> 2));
// output23Byte[19] = (byte) (((src[2] & 0x03) << 6) | src[3]);
// output23Byte[20] = (byte) ((src[4] << 2) | (src[5] >>> 4));
// output23Byte[21] = (byte) (((src[5] & 0x0F) << 4) | (src[6] >>> 2));
// output23Byte[22] = (byte) (((src[6] & 0x03) << 6) | src[7]);
// }
//
// /**
// * 32位加法: (a_hi:a_lo) + (b_hi:b_lo)
// * out[0] = lo, out[1] = hi
// */
// static void add32(short a_lo, short a_hi,
// short b_lo, short b_hi,
// short[] out /*len=2*/) {
//
// // ---- 低16位 ----
// short lo_low = (short)((a_lo & 0x00FF) + (b_lo & 0x00FF));
// short carry0 = (short)(((a_lo & 0x00FF) + (b_lo & 0x00FF)) >>> 8);
//
// short a_lo_hi = (short)((a_lo >>> 8) & 0x00FF);
// short b_lo_hi = (short)((b_lo >>> 8) & 0x00FF);
// short lo_high = (short)(a_lo_hi + b_lo_hi + carry0);
// short carry1 = (short)(lo_high >>> 8);
//
// short lo_res = (short)((lo_high << 8) | (lo_low & 0x00FF));
//
// // ---- 高16位 ----
// short hi_low = (short)((a_hi & 0x00FF) + (b_hi & 0x00FF) + carry1);
// short carry2 = (short)(hi_low >>> 8);
//
// short a_hi_hi = (short)((a_hi >>> 8) & 0x00FF);
// short b_hi_hi = (short)((b_hi >>> 8) & 0x00FF);
// short hi_high = (short)(a_hi_hi + b_hi_hi + carry2);
//
// short hi_res = (short)((hi_high << 8) | (hi_low & 0x00FF));
//
// // ---- 输出 ----
// out[0] = lo_res;
// out[1] = hi_res;
// }
//
// /**
// * 32位加法 + 返回进位（只用 short）
// * 输入: (a_hi:a_lo) + (b_hi:b_lo)
// * 输出: out[0]=lo, out[1]=hi
// * 返回: 最终进位(0/1)
// */
// static short add32_with_carry(short a_lo, short a_hi,
// short b_lo, short b_hi,
// short[] out /* len=2 */) {
// // ---- 低16位：分两段8位相加 ----
// short s0 = (short)((a_lo & (short)0x00FF) + (b_lo & (short)0x00FF)); // 0..510
// short c0 = (short)(s0 >>> 8); // 0/1
// short s1 = (short)(((a_lo >>> 8) & (short)0x00FF)
// + ((b_lo >>> 8) & (short)0x00FF)
// + c0); // 0..511
// short c1 = (short)(s1 >>> 8); // 0/1
// short lo = (short)((s1 << 8) | (s0 & (short)0x00FF));
//
// // ---- 高16位：再分两段8位相加，并加上 c1 ----
// short s2 = (short)((a_hi & (short)0x00FF) + (b_hi & (short)0x00FF) + c1);
// short c2 = (short)(s2 >>> 8); // 0/1
// short s3 = (short)(((a_hi >>> 8) & (short)0x00FF)
// + ((b_hi >>> 8) & (short)0x00FF)
// + c2); // 0..511
// short c3 = (short)(s3 >>> 8); // 最终进位 0/1
// short hi = (short)((s3 << 8) | (s2 & (short)0x00FF));
//
// out[0] = lo;
// out[1] = hi;
// return (short)(c3 & 1);
// }
//
//
//
// /**
// * 64位加法: a4 + b4 -> a4
// * 输入输出: short[4]，低到高 (a[0]=lo16, a[1]=hi16, a[2]=lo16 of high dword, a[3]=hi16 of high dword)
// */
// static void add64(short[] a, short[] b) {
// short[] tmp = JCSystem.makeTransientShortArray((short)2, JCSystem.MEMORY_TYPE_TRANSIENT_RESET);
//
// // 低 32 位
// short carry = add32_with_carry(a[0], a[1], b[0], b[1], tmp);
// a[0] = tmp[0];
// a[1] = tmp[1];
//
// // 高 32 位 + carry
// add32((short)(a[2] + (short)(carry & (short)0x0001)), a[3], b[2], b[3], tmp);
// a[2] = tmp[0];
// a[3] = tmp[1];
// }
//
//
//
// // 32位异或
// public static void xor32(short a_lo, short a_hi, short b_lo, short b_hi, short[] out /*len==2*/) {
// out[0] = (short)(a_lo ^ b_lo);
// out[1] = (short)(a_hi ^ b_hi);
// }
//
// /**
// * 把32位数 b (b[0]=lo, b[1]=hi) 左移 k 位 (0 <= k < 32)，
// * 结果放到64位数 a (a[0]=最低16位 ... a[3]=最高16位)。
// */
// static void create_64b_from_32b(short[] a/*len=4*/, short[] b/*len=2*/, short k) {
// short a0 = b[0], a1 = b[1], a2 = 0, a3 = 0;
//
// if (k >= 16) {
// a3 = a2; // 0
// a2 = a1; // 原 hi16
// a1 = a0; // 原 lo16
// a0 = 0;
// k = (short)(k - 16);
// }
//
// while (k > 0) {
// short c0 = (short)((a0 >>> 15) & 1);
// short c1 = (short)((a1 >>> 15) & 1);
// short c2 = (short)((a2 >>> 15) & 1);
//
// a3 = (short)((a3 << 1) | c2);
// a2 = (short)((a2 << 1) | c1);
// a1 = (short)((a1 << 1) | c0);
// a0 = (short)(a0 << 1);
// k--;
// }
//
// a[0] = a0; a[1] = a1; a[2] = a2; a[3] = a3;
// }
//
//
// /**
// * (A & 0x7FFFFFFF)，结果放在 out[4]，只保留低32位并清掉最高bit。
// */
// static void and64_7FFFFFFF_to32(short[] A, short[] out) {
// out[0] = A[0]; // lo16
// out[1] = (short)(A[1] & 0x7FFF); // hi16 (清除最高bit)
// out[2] = 0;
// out[3] = 0;
// }
//
// /**
// * 64位无符号右移 31 位
// * 输入: A[0..3] (short[4], A[0]最低16位)
// * 输出: out[0..3]
// */
// static void shr64u_31(short[] A, short[] out) {
// // 先拼出 64bit 的逻辑，逐段右移
// // A >>> 31 = (A >>> 16) >>> 15
//
// // 先右移 16，相当于丢掉 A[0]，整体右移一半字
// out[0] = A[1]; // 原 A[1] -> 新低16位
// out[1] = A[2]; // 原 A[2]
// out[2] = A[3]; // 原 A[3]
// out[3] = 0; // 高位补0
//
// // 再右移 15 位
// short c0 = (short)((out[0] & (short)0xFFFF) >>> 15); // out[0] 最后一位变进位
// short c1 = (short)((out[1] & (short)0xFFFF) >>> 15);
// short c2 = (short)((out[2] & (short)0xFFFF) >>> 15);
//
// out[0] = (short)((c0 & 0x0001) | (out[1] << 1));
// out[1] = (short)((c1 & 0x0001) | (out[2] << 1));
// out[2] = (short)(c2 & 0x0001);
// }
//
// /**
// * 32位无符号右移 1 位
// * 输入: lo,hi (short) 表示 32 位数 (hi:高16位, lo:低16位)
// * 输出: out[0]=lo, out[1]=hi
// */
// static void shr32u1(short lo, short hi, short[] out) {
// // >>>1：先处理低16位
// short nwLo = (short)(((((lo & (short)0xFFFF) >>> 1) & (short)0x7FFF)) | ((hi & 0x0001) << 15));
// short nwHi = (short)(((hi & (short)0xFFFF) >>> 1) & (short)0x7FFF);
//
// out[0] = nwLo;
// out[1] = nwHi;
// }
//
//
//
//
// /** 打印/*十六进制（调试用，TODO 生产/JC 环境可移除） *//*
// public static void printHex(String label, byte[] data, int len) {
// System.out.print(label + ": ");
// for (int i = 0; i < len; i++) {
// System.out.printf("%02x ", data[i] & 0xFF);
// }
// System.out.println();
// }*/
//}