On 11/21/2011 07:42 AM, Michael Matz wrote:
> __transaction_atomic
> {
> {
> struct shared_count sc;
> try
> // ctor(&sc), dtor(&sc)
> finally
> {
> sc = {CLOBBER};
> }
> }
> }
>
> later lowered to:
>
> <bb 3>:
> __transaction_atomic // SUBCODE=[ GTMA_HAVE_STORE ]
> <bb 4>:
> // ctor(&sc),dtor(&sc) [tm-clone]
> goto <bb 7>;
>
> <bb 7>:
> sc ={v} {CLOBBER};
> __builtin__ITM_commitTransaction ();
>
> <bb 8>:
> return 0;
>
> <L4>:
> landing-pad
>
> The object 'sc' is constructed on the stack. Is it necessary to make it
> transact? If not, then you can possibly improve code quality somewhat by
> ignoring the clobbers also in requires_barrier (which then would need
> changes to receive the whole statement, not just the lhs/rhs to determine
> if this is an ignorable statement). I haven't tried this, though. The
> patch below would be necessary anyway.
No, SC need not transact because it is already transaction-local.
> PR other/51125
> * trans-mem.c (expand_block_tm): Ignore clobbers.
Ok.
r~
