On Mon, May 5, 2014 at 8:57 AM, Richard Biener <richard.guent...@gmail.com> wrote: > On Mon, May 5, 2014 at 11:28 AM, Andrew Haley <a...@redhat.com> wrote: >> On 05/05/2014 08:47 AM, Richard Biener wrote: >>> It really depends on how "3x" should materialize in the end. >>> How do you triplicate ops with side-effects? If you only >>> triplicate ops without side-effects what is the sink that keeps >>> the duplicated ops live? >> >> The vote, surely. CSE would be absolutely determined to get rid of >> all this redundant work. I guess it'd all have to be done very late >> or we'd need a new node that CSE can't see through. >> >> Would it be enough to mark all nodes that go into the vote as volatile? >> I guess not: we wouldn't want to generate extra memory traffic. > > Well, it can end up using the exact same registers 3 times and spilling > the result to the stack anyway. > > It has to be done before register allocation I suppose, and thus > postreload-cse will come along ...
Thanks for all your feedback guys. Based on our experiments, we noted too many exposure in the stack frames, so it seems that a hardened ABI should be considered (with redundancy and voting too). All these facts lead to work on a specific RTL backend (e.g. hardened-x86, hardened-ARM, etc.). Any more comment will be greatly appreciated. Daniel. > > Richard. > >> Andrew. >> -- Daniel F. Gutson Chief Engineering Officer, SPD San Lorenzo 47, 3rd Floor, Office 5 Córdoba, Argentina Phone: +54 351 4217888 / +54 351 4218211 Skype: dgutson
