On 05/07/2013 11:20 AM, Han Shen(沈涵) wrote:
How do you plan on handling Florian's retslot issue? Are you going to scan
the gimple looking for suitable calls? How do you avoid instrumentation in
the callee for that case?
I find myself wondering if you'd be better off scanning the gimple
representation looking for ADDR_EXPR operations, then peeking at the
argument to see if it's part of the local frame or something higher up.
Prsumably you catch cases where the ABI mandates that certain arguments be
copied by the caller and the address passed to the callee? What about the
opposite (and yes, both exist). See TARGET_PASS_BY_REFERENCE,
TARGET_CALLEE_COPIES. I suspect scanning the gimple code is much more
likely to pick up this stuff.
Yeah, I see. That's are good suggestions. I'll amend these (probably
minor) cases in another patch - if no objection to that -
which needs a good testing/performance/security measurement.
OK for the trunk as long as you're agreeing to look at the retslot and
ABI mandated copies which might expose stack slots and report back to
this list.
I'll note you've got a couple merge markers (>>>>>) in your ChangeLogs.
Those need to be removed :-)
Thanks,
Jeff
