You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
CLs 21813 and 22197 optimized the code generated for append. It was done in part to address #14921, but while I was there, I also focused on things like avoiding spills in the fast path.
Several things have changed in the compiler since then:
The old backend is gone, so we are free to change the signature of growslice.
We are better at sinking spills to a narrower scope, so some of the spill-avoidance techniques used are no longer necessary.
Write barriers are inserted later in the SSA backend, rather than in the frontend and during SSA construction.
All this suggests to me that we may be able to simplify and improve the code generated for append.
I also suspect that the "no write barriers for in-place appends" optimization may have gotten lost during the write barrier move. More details about that in a comment-to-come over in #14921.
That's a significant optimization. Tentatively marking this as a release blocker for that reason.
CLs 21813 and 22197 optimized the code generated for append. It was done in part to address #14921, but while I was there, I also focused on things like avoiding spills in the fast path.
Several things have changed in the compiler since then:
All this suggests to me that we may be able to simplify and improve the code generated for append.
I also suspect that the "no write barriers for in-place appends" optimization may have gotten lost during the write barrier move. More details about that in a comment-to-come over in #14921.
That's a significant optimization. Tentatively marking this as a release blocker for that reason.
cc @randall77 @cherrymui @aclements @mvdan
The text was updated successfully, but these errors were encountered: