1  /* { dg-options "-O2" } */
       2  /* { dg-do run } */
       3  
       4  /* nvptx backend used to emit lo/ls/hs/hi suffixes on unsigned comparison
       5     insns instead of the more common lt/le/ge/gt, but ptxas and PTX JIT
       6     miscompile 'ls' and 'hi' under some circumstances, such as when the first
       7     source operand expands to a constant memory load, as demonstrated below.
       8     Reported as NVIDIA bug ID 1725195 (tracker is not public).  */
       9  
      10  /* Define this to observe PTX translation breakage.  */
      11  //#define EMIT_BROKEN_ASM 1
      12  
      13  /* Or define this to get expected codegen.  */
      14  //#define EMIT_WORKING_ASM 1
      15  
      16  static __attribute__((noinline,noclone)) int ls(unsigned a)
      17  {
      18      unsigned v;
      19      /* %nctaid.x is always 1 in gcc testing.  */
      20      asm ("mov.u32 %0, %%nctaid.x;" : "=r"(v));
      21  #if defined(EMIT_BROKEN_ASM)
      22      asm ("set.u32.ls.u32 %0, %1, %0;" : "+r"(a) : "r"(v));
      23  #elif defined(EMIT_WORKING_ASM)
      24      asm ("set.u32.le.u32 %0, %1, %0;" : "+r"(a) : "r"(v));
      25  #else
      26      a = v <= a ? -1 : 0;
      27  #endif
      28      return a;
      29  }
      30  static __attribute__((noinline,noclone)) int hi(unsigned a)
      31  {
      32      unsigned v;
      33      asm ("mov.u32 %0, %%nctaid.x;" : "=r"(v));
      34  #if defined(EMIT_BROKEN_ASM)
      35      asm ("set.u32.hi.u32 %0, %1, %0;" : "+r"(a) : "r"(v));
      36  #elif defined(EMIT_WORKING_ASM)
      37      asm ("set.u32.gt.u32 %0, %1, %0;" : "+r"(a) : "r"(v));
      38  #else
      39      a = v > a ? -1 : 0;
      40  #endif
      41      return a;
      42  }
      43  int main()
      44  {
      45      int i;
      46      for (i=0; i<3; i++)
      47          if (ls(i) != -(1 <= i) || hi(i) != -(1 > i))
      48              __builtin_abort();
      49      return 0;
      50  }