On Fri, 2011-10-28 at 07:53 -0500, Aldy Hernandez wrote:
> > diff --git a/gcc/testsuite/gcc.dg/tm/memopt-1.c
> > b/gcc/testsuite/gcc.dg/tm/memopt-1.c
> > index 06d4f64..9a48dcb 100644
> > --- a/gcc/testsuite/gcc.dg/tm/memopt-1.c
> > +++ b/gcc/testsuite/gcc.dg/tm/memopt-1.c
> > @@ -2,8 +2,8 @@
> > /* { dg-options "-fgnu-tm -O -fdump-tree-tmmemopt" } */
> >
> > long g, xxx, yyy;
> > -extern george() __attribute__((transaction_callable));
> > -extern ringo(long int);
> > +extern george() __attribute__((transaction_safe));
> > +extern ringo(long int) __attribute__((transaction_safe));
> > int i;
>
> The patch looks fine, but...
Looking closer at this, we were faking to have an uninstrumented code
path so not explicitly requesting irrevocable mode was okay. However, we
still had calls to the TM library in those code, which is not really
what we want (mostly for performance reasons, it is supposed to still
work because the runtime has to change to a suitable dispatch internally
after going irrevocable).
I'll prepare a different patch, after looking at Richard's recent
changes.
> Was the original test wrong, or are you testing something new?
Yes, sort of. It was testing for optimizations that were correct
performed but which should not have been applicable in this particular
test case _and_ with the current state of the code. However, after the
fix for when to request irrevocable mode that I have in mind, this test
should work as is.
Torvald
