libsnes/libco.patch

   1 --- a/libco/amd64.c     2011-04-30 12:18:27.987519764 +0200
   2 +++ b/libco/amd64.c     2011-04-30 12:19:04.959327913 +0200
   3 @@ -13,52 +13,94 @@
   4  extern "C" {
   5  #endif
   6
   7 +// ASM co_swap.
   8 +void co_swap(cothread_t new_active, cothread_t old_active);
   9 +
  10  static thread_local long long co_active_buffer[64];
  11  static thread_local cothread_t co_active_handle = 0;
  12 -static void (*co_swap)(cothread_t, cothread_t) = 0;
  13
  14  #ifdef _WIN32
  15 -  //ABI: Win64
  16 -  static unsigned char co_swap_function[] = {
  17 -    0x48, 0x89, 0x22, 0x48, 0x8B, 0x21, 0x58, 0x48, 0x89, 0x6A, 0x08, 0x48, 0x89, 0x72, 0x10, 0x48,
  18 -    0x89, 0x7A, 0x18, 0x48, 0x89, 0x5A, 0x20, 0x4C, 0x89, 0x62, 0x28, 0x4C, 0x89, 0x6A, 0x30, 0x4C,
  19 -    0x89, 0x72, 0x38, 0x4C, 0x89, 0x7A, 0x40, 0x48, 0x81, 0xC2, 0x80, 0x00, 0x00, 0x00, 0x48, 0x83,
  20 -    0xE2, 0xF0, 0x0F, 0x29, 0x32, 0x0F, 0x29, 0x7A, 0x10, 0x44, 0x0F, 0x29, 0x42, 0x20, 0x44, 0x0F,
  21 -    0x29, 0x4A, 0x30, 0x44, 0x0F, 0x29, 0x52, 0x40, 0x44, 0x0F, 0x29, 0x5A, 0x50, 0x44, 0x0F, 0x29,
  22 -    0x62, 0x60, 0x44, 0x0F, 0x29, 0x6A, 0x70, 0x44, 0x0F, 0x29, 0xB2, 0x80, 0x00, 0x00, 0x00, 0x44,
  23 -    0x0F, 0x29, 0xBA, 0x90, 0x00, 0x00, 0x00, 0x48, 0x8B, 0x69, 0x08, 0x48, 0x8B, 0x71, 0x10, 0x48,
  24 -    0x8B, 0x79, 0x18, 0x48, 0x8B, 0x59, 0x20, 0x4C, 0x8B, 0x61, 0x28, 0x4C, 0x8B, 0x69, 0x30, 0x4C,
  25 -    0x8B, 0x71, 0x38, 0x4C, 0x8B, 0x79, 0x40, 0x48, 0x81, 0xC1, 0x80, 0x00, 0x00, 0x00, 0x48, 0x83,
  26 -    0xE1, 0xF0, 0x0F, 0x29, 0x31, 0x0F, 0x29, 0x79, 0x10, 0x44, 0x0F, 0x29, 0x41, 0x20, 0x44, 0x0F,
  27 -    0x29, 0x49, 0x30, 0x44, 0x0F, 0x29, 0x51, 0x40, 0x44, 0x0F, 0x29, 0x59, 0x50, 0x44, 0x0F, 0x29,
  28 -    0x61, 0x60, 0x44, 0x0F, 0x29, 0x69, 0x70, 0x44, 0x0F, 0x29, 0xB1, 0x80, 0x00, 0x00, 0x00, 0x44,
  29 -    0x0F, 0x29, 0xB9, 0x90, 0x00, 0x00, 0x00, 0xFF, 0xE0,
  30 -  };
  31 -
  32 -  #include <windows.h>
  33 -
  34 -  void co_init() {
  35 -    DWORD old_privileges;
  36 -    VirtualProtect(co_swap_function, sizeof co_swap_function, PAGE_EXECUTE_READWRITE, &old_privileges);
  37 -  }
  38 +asm (
  39 +      ".text\n"
  40 +      ".globl co_swap\n"
  41 +      "co_swap:\n"
  42 +
  43 +      "movq %rsp, (%rdx) # Save stack pointer and pop old back\n"
  44 +      "movq (%rcx), %rsp\n"
  45 +
  46 +      "popq %rax # Pop saved PC to %rax\n"
  47 +
  48 +      "movq %rbp, 0x8(%rdx) # Save non-volatile registers\n"
  49 +      "movq %rsi, 0x10(%rdx)\n"
  50 +      "movq %rdi, 0x18(%rdx)\n"
  51 +      "movq %rbx, 0x20(%rdx)\n"
  52 +      "movq %r12, 0x28(%rdx)\n"
  53 +      "movq %r13, 0x30(%rdx)\n"
  54 +      "movq %r14, 0x38(%rdx)\n"
  55 +      "movq %r15, 0x40(%rdx)\n"
  56 +
  57 +      "addq $0x80, %rdx\n"
  58 +      "andq $0xfffffffffffffff0, %rdx # Align buffer to 16 bytes (SSE).\n"
  59 +      "movaps %xmm6, (%rdx) # Also save non-volatile vector registers.\n"
  60 +      "movaps %xmm7, 0x10(%rdx)\n"
  61 +      "movaps %xmm8, 0x20(%rdx)\n"
  62 +      "movaps %xmm9, 0x30(%rdx)\n"
  63 +      "movaps %xmm10, 0x40(%rdx)\n"
  64 +      "movaps %xmm11, 0x50(%rdx)\n"
  65 +      "movaps %xmm12, 0x60(%rdx)\n"
  66 +      "movaps %xmm13, 0x70(%rdx)\n"
  67 +      "movaps %xmm14, 0x80(%rdx)\n"
  68 +      "movaps %xmm15, 0x90(%rdx)\n"
  69 +
  70 +      "movq 0x8(%rcx),  %rbp # Pop back our saved registers and vector registers.\n"
  71 +      "movq 0x10(%rcx), %rsi\n"
  72 +      "movq 0x18(%rcx), %rdi\n"
  73 +      "movq 0x20(%rcx), %rbx\n"
  74 +      "movq 0x28(%rcx), %r12\n"
  75 +      "movq 0x30(%rcx), %r13\n"
  76 +      "movq 0x38(%rcx), %r14\n"
  77 +      "movq 0x40(%rcx), %r15\n"
  78 +
  79 +      "addq $0x80, %rcx\n"
  80 +      "andq $0xfffffffffffffff0, %rcx\n"
  81 +      "movaps %xmm6, (%rcx)\n"
  82 +      "movaps %xmm7, 0x10(%rcx)\n"
  83 +      "movaps %xmm8, 0x20(%rcx)\n"
  84 +      "movaps %xmm9, 0x30(%rcx)\n"
  85 +      "movaps %xmm10, 0x40(%rcx)\n"
  86 +      "movaps %xmm11, 0x50(%rcx)\n"
  87 +      "movaps %xmm12, 0x60(%rcx)\n"
  88 +      "movaps %xmm13, 0x70(%rcx)\n"
  89 +      "movaps %xmm14, 0x80(%rcx)\n"
  90 +      "movaps %xmm15, 0x90(%rcx)\n"
  91 +      "jmpq *%rax # Jump back to old saved PC.\n"
  92 +);
  93  #else
  94 -  //ABI: SystemV
  95 -  static unsigned char co_swap_function[] = {
  96 -    0x48, 0x89, 0x26, 0x48, 0x8B, 0x27, 0x58, 0x48, 0x89, 0x6E, 0x08, 0x48, 0x89, 0x5E, 0x10, 0x4C,
  97 -    0x89, 0x66, 0x18, 0x4C, 0x89, 0x6E, 0x20, 0x4C, 0x89, 0x76, 0x28, 0x4C, 0x89, 0x7E, 0x30, 0x48,
  98 -    0x8B, 0x6F, 0x08, 0x48, 0x8B, 0x5F, 0x10, 0x4C, 0x8B, 0x67, 0x18, 0x4C, 0x8B, 0x6F, 0x20, 0x4C,
  99 -    0x8B, 0x77, 0x28, 0x4C, 0x8B, 0x7F, 0x30, 0xFF, 0xE0,
 100 -  };
 101 -
 102 -  #include <unistd.h>
 103 -  #include <sys/mman.h>
 104 -
 105 -  void co_init() {
 106 -    unsigned long long addr = (unsigned long long)co_swap_function;
 107 -    unsigned long long base = addr - (addr % sysconf(_SC_PAGESIZE));
 108 -    unsigned long long size = (addr - base) + sizeof co_swap_function;
 109 -    mprotect((void*)base, size, PROT_READ | PROT_WRITE | PROT_EXEC);
 110 -  }
 111 +asm (
 112 +      ".text\n"
 113 +      ".globl co_swap\n"
 114 +      ".globl _co_swap\n"
 115 +      "_co_swap:\n"
 116 +      "co_swap:\n"
 117 +      "movq %rsp, (%rsi) # Save stack pointer, and pop the old one back.\n"
 118 +      "movq (%rdi), %rsp\n"
 119 +      "popq %rax\n"
 120 +
 121 +      "movq %rbp, 0x8(%rsi) # Save our non-volatile registers to buffer.\n"
 122 +      "movq %rbx, 0x10(%rsi)\n"
 123 +      "movq %r12, 0x18(%rsi)\n"
 124 +      "movq %r13, 0x20(%rsi)\n"
 125 +      "movq %r14, 0x28(%rsi)\n"
 126 +      "movq %r15, 0x30(%rsi)\n"
 127 +
 128 +      "movq 0x8(%rdi),  %rbp # Pop back our saved registers.\n"
 129 +      "movq 0x10(%rdi), %rbx\n"
 130 +      "movq 0x18(%rdi), %r12\n"
 131 +      "movq 0x20(%rdi), %r13\n"
 132 +      "movq 0x28(%rdi), %r14\n"
 133 +      "movq 0x30(%rdi), %r15\n"
 134 +      "jmpq *%rax # Jump back to saved PC.\n"
 135 +);
 136  #endif
 137
 138  static void crash() {
 139 @@ -72,15 +114,11 @@
 140
 141  cothread_t co_create(unsigned int size, void (*entrypoint)(void)) {
 142    cothread_t handle;
 143 -  if(!co_swap) {
 144 -    co_init();
 145 -    co_swap = (void (*)(cothread_t, cothread_t))co_swap_function;
 146 -  }
 147    if(!co_active_handle) co_active_handle = &co_active_buffer;
 148    size += 512; /* allocate additional space for storage */
 149    size &= ~15; /* align stack to 16-byte boundary */
 150
 151 -  if(handle = (cothread_t)malloc(size)) {
 152 +  if((handle = (cothread_t)malloc(size))) {
 153      long long *p = (long long*)((char*)handle + size); /* seek to top of stack */
 154      *--p = (long long)crash;                           /* crash if entrypoint returns */
 155      *--p = (long long)entrypoint;                      /* start of function */