This is my first stab at disallowing load data races that happen when we 
cache the value of a global.  I would appreciate input before 
committing, especially on the RTL bits, cause it's been quite a while 
since I typed d-e-b-u-g-_-r-t-x. :-)

In the example below we usually hoist "global" into a register or 
temporary to avoid reading from it at each step.  This would cause a 
race if another thread had modified "global" in between iterations.

	for (x=0; x< 5; x++)
		sum[x] =  global;

As expected, multiple passes can cause the same end result, so plugging 
the problem involves multiple places.  First, loop invariant movement 
moves the invariant.  Barring that, PRE moves the load, and even if we 
get past the SSA passes, rtl-CSE pulls the rug from under us.  If that 
weren't enough, the post-reload pass performs CSE over the hard 
registers, and we end up eliminating subsequent loads of "global".  I am 
sure there are many other places, but I'm starting with whatever it 
takes to fix gcc.dg/memmodel/global-hoist.c.

The patch below fixes the test in question with "--param 
allow-load-data-races=0".  What do y'all think?

Thanks.