Update: 睡前简单弄个小更新 把我回文中说的code cache的hashtable快速弄出来了 第一次呼叫时会emite code，第二次因为hashtable查找成功则直接呼叫 另外就是把mmap要记忆体时的可读可写可执行，改成仅可读可写 後面再透过mprotect改成可读可执行 这样可以防止後来有人窜改JIT emit的code cache patch放在这边： https://paste.plurk.com/show/2636452/ P.S. 我不用github放的原因是因为我不想暴露我的主ID 还请担待 = = = = = = = = = = = = 中午吃饱饭有点胀气 看到某篇回文的推文中有人想要看JIT范例 所以简单写了个很粗糙的版本 前後花不到15分钟 所以觉得coding style很烂、没效率是很正常的，别打我QQ (结果真的被转出去了Orz 我晚点忙完会来修code 真的assembly code gen会找时间补上的） 大抵上的思路就是你吃到虚拟机的bytecode， 就把它转成host平台的native code 接着把它塞进你跟系统动态要的、可以执行的记忆体区段 然後就很快乐的开始执行它 黄色上色的地方，每个作业系统给的API不一样 Linux、OSX、FreeBSD ... (most of *NIX OSes)： mprotect()、mmap() with proper permissions. Windows： VirtualAlloc() = = = = = = = = = = = /** * * This piece of code is to demo simple VM/JIT knowledge * and is released under 2-clause BSD license. * * Copyright 2018/05/16 snaketsai * Redistribution and use in source and binary forms, with or without modification, * are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * U2FsdGVkX19w8Ikk1T7xBlbh4vDhIEvZzshUhXft6XMFugC9M27uV9LDszf7/8gP * OtF2AZwYaUQqzLLY5vXhCQ== * **/ #include <stdio.h> #include <stdlib.h> #include <string.h> #include <sys/mman.h> #ifdef _NOJIT static inline int IAdd(int inp1, int inp2) { puts("Non-JIT version iadd."); return inp1+inp2; } static inline int ISub(int inp1, int inp2) { puts("Non-JIT version isub."); return inp1-inp2; } #else // Nope, I'm not gonna write an assembler. // Just pretend that we _magically_ get assembly pieces we need. const unsigned char _add[] = \ "\x55\x48\x89\xe5\x89\x7d\xfc\x89\x75\xf8\x8b\x55\xfc\x8b\x45\xf8\x01\xd0\x5d\xc3\x00"; const unsigned char _sub[] = \ "\x55\x48\x89\xe5\x89\x7d\xfc\x89\x75\xf8\x8b\x45\xfc\x2b\x45\xf8\x5d\xc3\x00"; #define MAXCODECAHE_SIZE 4096 void* AllocExeMem(size_t size) { void* memPtr = mmap(0, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); if (memPtr == (void*)-1) { perror("mmap() failed."); return NULL; } return memPtr; } void *CodeCachePool; int (*IAdd)(int, int); int (*ISub)(int, int); #endif /** Very Simple VM spec - - Every insn. is 3 byte long. opcodes are listed below: 0x00 HCF, halt and catch fire. 0x01 IADD, humbly add two signed integers and print the result. 0x02 ISUB, humbly substract two signed integers and print the result. The bytecode stream should end with a 0xff byte mark. **/ const unsigned char BytecodeStream[] = \ "\x01\x02\x03\x02\x05\x04\x00\x00\x00\xff"; int runVM(const unsigned char* bstream) { unsigned char insn[3]; while(bstream[0] != 0xff) { memcpy(&insn,bstream,3); switch(insn[0]) { case 0x00: puts("Dave, stop, will you ?"); return 0; case 0x01: #ifndef _NOJIT // emit code to code cache. memcpy(CodeCachePool, _add, sizeof(_add)); #endif printf("iadd: %d\n", IAdd((int)insn[1], (int)insn[2])); break; case 0x02: #ifndef _NOJIT memcpy(CodeCachePool, _sub, sizeof(_sub)); #endif printf("isub: %d\n", ISub((int)insn[1], (int)insn[2])); break; default : // Unrecognized insn. perror("Sorry Dave, I can't do that."); return -1; } bstream+=3; } } int main(int ac, char* av[]) { int ret = -1; #ifndef _NOJIT CodeCachePool = AllocExeMem(MAXCODECAHE_SIZE); IAdd = CodeCachePool; ISub= CodeCachePool; #endif ret = runVM(BytecodeStream); if(ret == 0) { perror("VM exited successfully."); } else { perror("Unexpected error occured."); } return 0; } -- Linux is the bone of my world. Kernel is mybody, and initramfs is myblood, have created over a thousand Distros. Unknown to impossibility. Nor known to limitation. --

※ 发信站: 批踢踢实业坊(ptt.cc), 来自: 140.113.167.181 ※ 文章网址: https://webptt.com/cn.aspx?n=bbs/Soft_Job/M.1526451691.A.2E2.html ※ 编辑: snaketsai (140.113.167.181), 05/16/2018 14:58:35 这整份code其实没什麽技术背景，大概上过compiler课的人都会写 唯一需要额外知道的，是如何去要一块可塞code又可执行的记忆体 这个课堂上不会教 没欸，会那麽长是因为包含function call要的prologue跟epilogue 我其实弄得比较像intrinsic fnuction 所以Non-JIT的部份可以直接跟JIT的无缝接在一起 没想到有这麽完整的write up 马上google一下"mprotect/mmap"+"JIT"发现已经不少文章在讲 我第一次知道JIT大概是2011年，当时文章还真的不多 稍微说一下我这篇目前离实用code的差距在： (1) 缺真正的asm codegen，我上面的asm snippet是用gcc -c去编C code来生的 (2) JIT emit code後没有code reuse、每次都还会重新emit 实务上会是个hashtable，查找失败emit、查找成功直接call 更进阶一点会有所谓的block chaining，也就是把能接在一起的native code串在一起跑 这样可以节省跳进、又跳出VM code的时间 (3) mmap权限直接开executable其实有点危险，比较正确的作法会分两段： 先mmap开rw要记忆体、把code塞进去 然後用mprotect把它改设成rx ※ 编辑: snaketsai (140.113.167.181), 05/16/2018 23:24:51 应该是那样没错 ※ 编辑: snaketsai (140.113.167.181), 05/17/2018 02:53:13