VEX/test/fxsave.c

   1
   2 #include <stdio.h>
   3 #include <stdlib.h>
   4 #include <malloc.h>
   5 #include <string.h>
   6
   7 const unsigned int vec0[4]
   8    = { 0x12345678, 0x11223344, 0x55667788, 0x87654321 };
   9
  10 const unsigned int vec1[4]
  11    = { 0xABCDEF01, 0xAABBCCDD, 0xEEFF0011, 0x10FEDCBA };
  12
  13 const unsigned int vecZ[4]
  14    = { 0, 0, 0, 0 };
  15
  16 void do_fxsave ( void* p ) {
  17    asm __volatile__("fxsave (%0)" : : "r" (p) : "memory" );
  18 }
  19
  20 void do_fxrstor ( void* p ) {
  21    asm __volatile__("fxrstor (%0)" : : "r" (p) : "memory" );
  22 }
  23
  24 void do_zeroise ( void )
  25 {
  26    asm __volatile__("finit");
  27    asm __volatile__(
  28     "fldz\n\t"
  29     "fldz\n\t"
  30     "fldz\n\t"
  31     "fldz\n\t"
  32     "fldz\n\t"
  33     "fldz\n\t"
  34     "fldz\n\t"
  35     "fldz\n\t"
  36     "finit\n");
  37    asm __volatile__("movups vecZ, %xmm0");
  38    asm __volatile__("movups vecZ, %xmm1");
  39    asm __volatile__("movups vecZ, %xmm2");
  40    asm __volatile__("movups vecZ, %xmm3");
  41    asm __volatile__("movups vecZ, %xmm4");
  42    asm __volatile__("movups vecZ, %xmm5");
  43    asm __volatile__("movups vecZ, %xmm6");
  44    asm __volatile__("movups vecZ, %xmm7");
  45    asm __volatile__(
  46       "pushl $0\n\t"
  47       "ldmxcsr 0(%esp)\n\t"
  48       "addl $4,%esp\n");
  49 }
  50
  51 /* set up the FP and SSE state, and then dump it. */
  52 void do_setup_then_fxsave ( void* p )
  53 {
  54    asm __volatile__("finit");
  55    asm __volatile__("fldpi");
  56    asm __volatile__("fld1");
  57    asm __volatile__("fldln2");
  58    asm __volatile__("fldlg2");
  59    asm __volatile__("fld %st(3)");
  60    asm __volatile__("fld %st(3)");
  61    asm __volatile__("movups vec0, %xmm0");
  62    asm __volatile__("movups vec1, %xmm1");
  63    asm __volatile__("xorps %xmm2, %xmm2");
  64    asm __volatile__("movaps %xmm2, %xmm3");
  65    asm __volatile__("movaps %xmm2, %xmm4");
  66    asm __volatile__("movaps %xmm2, %xmm5");
  67    asm __volatile__("movaps %xmm2, %xmm6");
  68    asm __volatile__("movaps %xmm1, %xmm7");
  69    asm __volatile__("xorps %xmm0, %xmm7");
  70    do_fxsave (p);
  71 }
  72
  73 int isFPLsbs ( int i )
  74 {
  75    int q;
  76    q = 32; if (i == q || i == q+1) return 1;
  77    q = 48; if (i == q || i == q+1) return 1;
  78    q = 64; if (i == q || i == q+1) return 1;
  79    q = 80; if (i == q || i == q+1) return 1;
  80    q = 96; if (i == q || i == q+1) return 1;
  81    q = 112; if (i == q || i == q+1) return 1;
  82    q = 128; if (i == q || i == q+1) return 1;
  83    q = 144; if (i == q || i == q+1) return 1;
  84    return 0;
  85 }
  86
  87 void show ( unsigned char* buf, int xx )
  88 {
  89    int i;
  90    for (i = 0; i < 512; i++) {
  91       if ((i % 16) == 0)
  92          printf("%3d   ", i);
  93       if (xx && isFPLsbs(i))
  94          printf("xx ");
  95       else
  96          printf("%02x ", buf[i]);
  97       if (i > 0 && ((i % 16) == 15))
  98           printf("\n");
  99    }
 100 }
 101
 102
 103 int main ( int argc, char** argv )
 104 {
 105    unsigned char* buf1 = memalign(16,512);
 106    unsigned char* buf2 = memalign(16,512);
 107    unsigned char* buf3 = memalign(16,512);
 108    int xx = argc > 1;
 109    printf("Re-run with any arg to suppress least-significant\n"
 110           "   16 bits of FP numbers\n");
 111    memset(buf1, 0x55, 512);
 112    memset(buf2, 0x55, 512);
 113    memset(buf3, 0x55, 512);
 114
 115    /* Load up x87/xmm state and dump it. */
 116    do_setup_then_fxsave(buf1);
 117    printf("\nBEFORE\n");
 118    show(buf1, xx);
 119
 120    /* Zeroise x87/xmm state and dump it, to show that the
 121       regs have been cleared out. */
 122    do_zeroise();
 123    do_fxsave(buf2);
 124    printf("\nZEROED\n");
 125    show(buf2, xx);
 126
 127    /* Reload x87/xmm state from buf1 and dump it in buf3. */
 128    do_fxrstor(buf1);
 129    do_fxsave(buf3);
 130    printf("\nRESTORED\n");
 131    show(buf3, xx);
 132
 133    free(buf1); free(buf2); free(buf3);
 134
 135    return 0;
 136 }