memcheck/tests/x86/fxsave.c

   1
   2 #include <stdio.h>
   3 #include <stdlib.h>
   4 #include "tests/asm.h"
   5 #include "tests/malloc.h"
   6 #include <string.h>
   7
   8 const unsigned int vec0[4]
   9    = { 0x12345678, 0x11223344, 0x55667788, 0x87654321 };
  10
  11 const unsigned int vec1[4]
  12    = { 0xABCDEF01, 0xAABBCCDD, 0xEEFF0011, 0x10FEDCBA };
  13
  14 const unsigned int vecZ[4]
  15    = { 0, 0, 0, 0 };
  16
  17 void do_fxsave ( void* p ) {
  18    asm __volatile__("fxsave (%0)" : : "r" (p) : "memory" );
  19 }
  20
  21 void do_fxrstor ( void* p ) {
  22    asm __volatile__("fxrstor (%0)" : : "r" (p) : "memory" );
  23 }
  24
  25 void do_zeroise ( void )
  26 {
  27    asm __volatile__("finit");
  28    asm __volatile__(
  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     "fldz\n\t"
  37     "finit\n");
  38    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm0");
  39    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm1");
  40    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm2");
  41    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm3");
  42    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm4");
  43    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm5");
  44    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm6");
  45    asm __volatile__("movups " VG_SYM(vecZ) ", %xmm7");
  46    asm __volatile__(
  47       "pushl $0\n\t"
  48       "ldmxcsr 0(%esp)\n\t"
  49       "addl $4,%esp\n");
  50 }
  51
  52 /* set up the FP and SSE state, and then dump it. */
  53 void do_setup_then_fxsave ( void* p )
  54 {
  55    asm __volatile__("finit");
  56    asm __volatile__("fldpi");
  57    asm __volatile__("fld1");
  58    asm __volatile__("fldln2");
  59    asm __volatile__("fldlg2");
  60    asm __volatile__("fld %st(3)");
  61    asm __volatile__("fld %st(3)");
  62    asm __volatile__("movups " VG_SYM(vec0) ", %xmm0");
  63    asm __volatile__("movups " VG_SYM(vec1) ", %xmm1");
  64    asm __volatile__("xorps %xmm2, %xmm2");
  65    asm __volatile__("movaps %xmm2, %xmm3");
  66    asm __volatile__("movaps %xmm2, %xmm4");
  67    asm __volatile__("movaps %xmm2, %xmm5");
  68    asm __volatile__("movaps %xmm2, %xmm6");
  69    asm __volatile__("movaps %xmm1, %xmm7");
  70    asm __volatile__("xorps %xmm0, %xmm7");
  71    do_fxsave (p);
  72 }
  73
  74 int isFPLsbs ( int i )
  75 {
  76    int q;
  77    q = 32; if (i == q || i == q+1) return 1;
  78    q = 48; if (i == q || i == q+1) return 1;
  79    q = 64; if (i == q || i == q+1) return 1;
  80    q = 80; if (i == q || i == q+1) return 1;
  81    q = 96; if (i == q || i == q+1) return 1;
  82    q = 112; if (i == q || i == q+1) return 1;
  83    q = 128; if (i == q || i == q+1) return 1;
  84    q = 144; if (i == q || i == q+1) return 1;
  85    return 0;
  86 }
  87
  88 void show ( unsigned char* buf, int xx )
  89 {
  90    int i;
  91    for (i = 0; i < 512; i++) {
  92       if ((i % 16) == 0)
  93          printf("%3d   ", i);
  94       if (xx && isFPLsbs(i))
  95          printf("xx ");
  96       else
  97          printf("%02x ", buf[i]);
  98       if (i > 0 && ((i % 16) == 15))
  99           printf("\n");
 100    }
 101 }
 102
 103
 104 int main ( int argc, char** argv )
 105 {
 106    unsigned char* buf1 = memalign16(512);
 107    unsigned char* buf2 = memalign16(512);
 108    unsigned char* buf3 = memalign16(512);
 109    int xx = argc > 1;
 110    printf("Re-run with any arg to suppress least-significant\n"
 111           "   16 bits of FP numbers\n");
 112    memset(buf1, 0x55, 512);
 113    memset(buf2, 0x55, 512);
 114    memset(buf3, 0x55, 512);
 115
 116    /* Load up x87/xmm state and dump it. */
 117    do_setup_then_fxsave(buf1);
 118    printf("\nBEFORE\n");
 119    show(buf1, xx);
 120
 121    /* Zeroise x87/xmm state and dump it, to show that the
 122       regs have been cleared out. */
 123    do_zeroise();
 124    do_fxsave(buf2);
 125    printf("\nZEROED\n");
 126    show(buf2, xx);
 127
 128    /* Reload x87/xmm state from buf1 and dump it in buf3. */
 129    do_fxrstor(buf1);
 130    do_fxsave(buf3);
 131    printf("\nRESTORED\n");
 132    show(buf3, xx);
 133
 134    free(buf1); free(buf2); free(buf3);
 135
 136    return 0;
 137 }