[Bug target/67281] HTM builtins aren't treated as compiler barriers on powerpc

[tuliom at linux dot vnet.ibm.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67281

--- Comment #4 from Tulio Magno Quites Machado Filho <tuliom at linux dot vnet.ibm.com> ---
(In reply to Andrew Pinski from comment #3)
> Since there are no stores, the load seems like it can be pulled out of the
> loop too.

I disagree with you.
If I use the value of dest to take a decision inside the transaction, I need
the memory barrier before the access to *src0.

Here's an example:

long
foo (long dest, long *src0, long src1, long tries)
{
  long i;
  for (i = 0; i < tries; i++)
    {
      __builtin_tbegin (0);
      dest = *src0 + src1;
      if (dest == 13)
        __builtin_tabort(0);
      __builtin_tend (0);
    }
  return dest;
}

In other words, if you access *src0 before the memory barrier, its value may
change when the memory barrier is created.  This is particularly useful if dest
says if a lock has been acquired by another thread or not.

For the reference, this has been found in glibc source code:
https://sourceware.org/ml/libc-alpha/2015-07/msg00986.html