[turbofan]: Elide extra move when accessing stack or frame register

[turbofan]: Elide extra move when accessing stack or frame register

Before this CL, the kFramePointer and kStackPointer IR instructions in turbo fan
moved the values of the frame and stack pointers into explicitly allocated
temporary registers. With this change, the instruction selector and code
generator now handle the stack and frame register specially, allowing them to be
directly used by generated code without an intermediate move.

[danno, 2015-07-22 08:46:07]

danno@chromium.org changed reviewers:
+ jarin@chromium.org, titzer@chromium.org

[danno, 2015-07-22 08:46:07]

PTAL

[Jarin, 2015-07-23 08:13:01]

lgtm.

I am wondering whether the extra complexity is going to pay for itself. The
special casing OperandGenerator::Use looks funky (but I do not know of a better
way to do this).

[Benedikt Meurer, 2015-07-24 06:02:46]

bmeurer@chromium.org changed reviewers:
+ bmeurer@chromium.org

[Benedikt Meurer, 2015-07-24 06:02:46]

NOT LGTM. This change is not correct, because the stack pointer is not constant
across a function (at the instruction level). We currently have explicit Push
operations there, and we might have other things in the future. Plus we are
still not sure whether LoadStackPointer should also be effectful.

Personally I'd also consider this change bad wrt. product excellence strategy,
because it's a lot of complexity for no real benefit (I'd guess that even in the
hottest paths, this only gives you up to 5% if at all).

[danno, 2015-07-24 06:45:34]

On 2015/07/24 at 06:02:46, bmeurer wrote:
> NOT LGTM. This change is not correct, because the stack pointer is not
constant across a function (at the instruction level). We currently have
explicit Push operations there, and we might have other things in the future.
Plus we are still not sure whether LoadStackPointer should also be effectful.
> 
> Personally I'd also consider this change bad wrt. product excellence strategy,
because it's a lot of complexity for no real benefit (I'd guess that even in the
hottest paths, this only gives you up to 5% if at all).

Fair enough. It was meant to flush out a discussion that Jaro and I had earlier
this week. I missed the incorrectness because of the Push, and I can understand
the argument about the complexity.