It has come to my attention, by..ahem...you, that in handling the 
testcase from my previous patch:

   __transaction_relaxed { __asm__(""); }

...if we avoid instrumentation altogether, we shouldn't lie to the 
run-time and pass PR_INSTRUMENTEDCODE.

I suppose we could cheat and avoid passing PR_INSTRUMENTEDCODE if we 
ever enter expand_block_tm(), but perhaps we could do a little better as 
with the attached patch.

With this patch, we generate:

   tm_state.2_6 = __builtin__ITM_beginTransaction (16458); [ 
uninstrumentedCode hasNoAbort doesGoIrrevocable readOnly ]

Notice there is no "instrumentedCode".

By the way, can you double check my logic in expand_assign_tm() and 
expand_call_tm(), particularly if we should bother setting 
GTMA_INSTRUMENTED_CODE for thread private variables?

Thanks.