|
| 1 | +#include <cerrno> |
| 2 | +#include <cstdint> |
| 3 | +#include <fstream> |
| 4 | +#include <filesystem> |
| 5 | +#include <iostream> |
| 6 | +#include <libkern/OSCacheControl.h> |
| 7 | +#include <stack> |
| 8 | +#include <stdexcept> |
| 9 | +#include <sys/mman.h> |
| 10 | + |
| 11 | +// The number of bytes an instruction is wide. AArch64 is 32-bit. |
| 12 | +#define INS_WIDTH 4 |
| 13 | + |
| 14 | +using std::uintmax_t; |
| 15 | +using std::fstream; |
| 16 | +using std::stack; |
| 17 | + |
| 18 | +// We work with a 50k memory cell, could be more, could be less. |
| 19 | +// Wrap-around is forbidden, so it's good to have a large space. |
| 20 | +constexpr size_t MEMORY_SIZE = 50000; |
| 21 | + |
| 22 | +// Compile a single instruction. |
| 23 | +// This will take advantage of the compilation state, and if applicable |
| 24 | +// patch previous compilations. |
| 25 | +void compile(uint32_t* baseAddress, |
| 26 | + size_t &pc, |
| 27 | + char &c, |
| 28 | + stack<size_t> &jumpPoints); |
| 29 | + |
| 30 | +// Creates an executable memory region sufficiently large to contain all the |
| 31 | +// instructions in the program. |
| 32 | +uint32_t* createExecutableMemory(uintmax_t length); |
| 33 | + |
| 34 | +// Sets up the stack/registers. |
| 35 | +void prelude(uint32_t* baseAddress, size_t &pc, char (&memory)[MEMORY_SIZE]); |
| 36 | + |
| 37 | +// Restores the stack and returns execution. |
| 38 | +void postlude(uint32_t* baseAddress, uintmax_t instructionSize, size_t &pc); |
| 39 | + |
| 40 | +// Helper function to write a 32-bit instruction to the byte array. |
| 41 | +void writeInstruction(uint32_t* baseAddress, size_t &pc, uint32_t instruction); |
| 42 | + |
| 43 | +// The main function takes in arguments and then executes the code. |
| 44 | +int main(int argc, char** argv) { |
| 45 | + // We expect most of the time the program is provided. |
| 46 | + if (__builtin_expect(argc < 2, 0)) { |
| 47 | + std::cerr << "zero: please provide the input file" << std::endl; |
| 48 | + return 1; |
| 49 | + } |
| 50 | + // Get the file name, and consequently read the size of the file. |
| 51 | + // In the JIT version, we do not accept non-code characters in order to be |
| 52 | + // able to very quickly estimate the instruction file size. |
| 53 | + char* fileName = argv[1]; |
| 54 | + uintmax_t fileSize = std::filesystem::file_size(fileName); |
| 55 | + // Perform the memory mapping. This is a pessimistic estimate. We assume that |
| 56 | + // we perform 4 instructions (128 bits) per BF instruction, plus an additional |
| 57 | + // 8 instructions for setup and teardown. |
| 58 | + uintmax_t instructionSize = (INS_WIDTH * 4 * fileSize) + (INS_WIDTH * 8); |
| 59 | + uint32_t* baseAddress = createExecutableMemory(instructionSize); |
| 60 | + // Keep track of the program address and memory. |
| 61 | + size_t pc = 0; |
| 62 | + char memory[MEMORY_SIZE] = {0}; |
| 63 | + // Write the assembly prelude. |
| 64 | + prelude(baseAddress, pc, memory); |
| 65 | + // Keep track of the jump points for loops. |
| 66 | + stack<size_t> jumpPoints; |
| 67 | + char ch; |
| 68 | + // Read the file line by line and write it to the address space. |
| 69 | + fstream file(fileName, fstream::in); |
| 70 | + while (file >> ch) { |
| 71 | + compile(baseAddress, pc, ch, jumpPoints); |
| 72 | + // Offset by one instruction. |
| 73 | + //pc += INS_WIDTH; |
| 74 | + } |
| 75 | + if (!__builtin_expect(jumpPoints.empty(), 1)) { |
| 76 | + throw std::runtime_error("program parse error: expected ]"); |
| 77 | + } |
| 78 | + // Write the assembly postlude. |
| 79 | + postlude(baseAddress, instructionSize, pc); |
| 80 | + // Jump to the actual JIT subroutine. |
| 81 | + return reinterpret_cast<uint32_t(*)()>(baseAddress)(); |
| 82 | +} |
| 83 | + |
| 84 | +// Actually perform the compilation. This should modify the program counter. |
| 85 | +// The register allocation is as follows: |
| 86 | +// x19 -- base address of the memory cells [callee saved]. |
| 87 | +// x20 -- memory index [callee saved]. |
| 88 | +inline void compile(uint32_t* baseAddress, |
| 89 | + size_t &pc, |
| 90 | + char &c, |
| 91 | + stack<size_t> &jumpPoints) { |
| 92 | + //std::cout << c << std::endl; |
| 93 | +} |
| 94 | + |
| 95 | +// The kernel will automatically unmap this in program exit. |
| 96 | +// This positively influences the runtime of the compiled binary. |
| 97 | +inline uint32_t* createExecutableMemory(uintmax_t length) { |
| 98 | + void* startingAddress = mmap(nullptr, |
| 99 | + length, |
| 100 | + PROT_READ | PROT_WRITE | PROT_EXEC, |
| 101 | + MAP_PRIVATE | MAP_ANON | MAP_JIT, |
| 102 | + -1, |
| 103 | + 0); |
| 104 | + return reinterpret_cast<uint32_t*>(startingAddress); |
| 105 | +} |
| 106 | + |
| 107 | +inline void prelude(uint32_t* baseAddress, |
| 108 | + size_t &pc, |
| 109 | + char (&memory)[MEMORY_SIZE]) { |
| 110 | + // Need to give permission to write to the memory. |
| 111 | + pthread_jit_write_protect_np(0); |
| 112 | + writeInstruction(baseAddress, pc, 0b11010110010111110000001111000000); |
| 113 | + return; |
| 114 | + // Start by snapshotting callee saved registers to the stack: |
| 115 | + // stp x19, x20, [sp, #-16]! |
| 116 | + writeInstruction(baseAddress, pc, 0xf353bfa9); |
| 117 | + // mov x20, #0 |
| 118 | + writeInstruction(baseAddress, pc, 0x140080d2); |
| 119 | + // mov x0, xzr |
| 120 | + writeInstruction(baseAddress, pc, 0xe0031faa); |
| 121 | +} |
| 122 | + |
| 123 | +inline void postlude(uint32_t* baseAddress, |
| 124 | + uintmax_t instructionSize, |
| 125 | + size_t &pc) { |
| 126 | + // Restore the callee saved registers. |
| 127 | + // ldp x19, x20, [sp], #16 |
| 128 | + //writeInstruction(baseAddress, pc, 0xf353c1a8); |
| 129 | + // ret |
| 130 | + //writeInstruction(baseAddress, pc, 0xc0035fd6); |
| 131 | + // Disallow writing to the memory region again. |
| 132 | + pthread_jit_write_protect_np(1); |
| 133 | + // Flush the instruction cache explicitly. |
| 134 | + sys_icache_invalidate(baseAddress, instructionSize); |
| 135 | + char* charAddress = reinterpret_cast<char*>(baseAddress); |
| 136 | + __builtin___clear_cache(charAddress, charAddress + instructionSize); |
| 137 | +} |
| 138 | + |
| 139 | +inline void writeInstruction(uint32_t* baseAddress, |
| 140 | + size_t &pc, |
| 141 | + uint32_t instruction) { |
| 142 | + baseAddress[pc] = instruction; |
| 143 | + pc++; |
| 144 | +} |
0 commit comments