[LLVMdev] correctness in localized global variable between setjmp and longjmp

[Rafael Espíndola]

On 28 April 2014 15:01, Balaram Makam <bmakam at codeaurora.org> wrote:
> I noticed that c code below failed with -O3 :
>
>
>
> #include <setjmp.h>
>
> #include <stdio.h>
>
>
>
> static int cnt;
>
>
>
> int main(){
>
>   int var;
>
>   jmp_buf buffer;
>
>   cnt = 0;
>
>
>
>   if ((var = setjmp(buffer)) == 0) {
>
>     printf(" if true, var: %d, .count:%d , increase count !\n",var, cnt);
>
>     cnt++;
>
>     longjmp(buffer, 2);
>
>   } else {
>
>     if (cnt == 1)
>
>       printf(" Pass      var: %d, .count:%d \n",var, cnt);
>
>     else
>
>       printf(" Fail !!!  var: %d, .count:%d \n",var, cnt);
>
>   }
>
>   return 0;
>
> }
>
>
>
> The documentation of setjmp/longjmp says that the values of objects of
> automatic storage duration which are local to the function containing the
> invocation of the corresponding setjmp() which do not have
> volatile-qualified type and which are changed between the setjmp()
> invocation and longjmp() call are indeterminate.
>
>
>
> Even though “cnt” is a global variable, above c code failed because the
> globalopt pass localizes the global value (cnt) in main(), causing that the
> value of the global variable becomes indeterminate between setjmp() and
> longjmp. I believe the globalopt pass need to check this case before
> localizing global variables. Is there any better solution about it?

No, looks like it has to check it.

This is a very specific optimization, but it looks like it is still
important in old benchmarks.

Cheers,
Rafael