On 2012-11-01 07:31, Aldy Hernandez wrote:
> + // Hmmm, the front-end should have caught outer aborts without
> + // an outer transaction. Bail and hope for the best.
> + tree attrs = get_attrs_for (current_function_decl);
> + if (!attrs || !lookup_attribute ("transaction_may_cancel_outer",
> + attrs))
> + return;
No, like
// The front end should have semantically checked outer aborts, but in either
case
// the target region is not within this function.
gcc_checking_assert(attrs && lookup_attribute
("transaction_may_cancel_outer", attrs));
return;
And now that you mention it...
> - if (subcode & GTMA_DOES_GO_IRREVOCABLE)
> - flags = PR_DOESGOIRREVOCABLE | PR_UNINSTRUMENTEDCODE;
> - else
> - flags = PR_INSTRUMENTEDCODE;
...
> + int flags = PR_MULTIWAYCODE;
> + tree tm_start = builtin_decl_explicit (BUILT_IN_TM_START);
> +
> + /* ??? There are plenty of bits here we're not computing. */
> + int subcode = gimple_transaction_subcode (region->transaction_stmt);
> + if (subcode & GTMA_DOES_GO_IRREVOCABLE)
> + flags |= PR_DOESGOIRREVOCABLE;
This is a regression for GTMA_DOES_GO_IRREVOCABLE. We don't want to
generate instrumented code in that case. We certainly didn't before.
And lets not advertise uninstrumented code until we've got it.
r~
