As discussed in the recent reduction speedup work:
https://github.com/openjdk/jdk/pull/13056JDK-8302652
We shoud intrinsify it, and keep a MaxL / MinL node in the graph.
In the backend, we can then use instructions, or revert back to
a compare and jump / CMoveL.
Benefit: SuperWord could vectorize MaxL/MinL. Probably generally leads to speedups.
https://github.com/openjdk/jdk/pull/13056
We shoud intrinsify it, and keep a MaxL / MinL node in the graph.
In the backend, we can then use instructions, or revert back to
a compare and jump / CMoveL.
Benefit: SuperWord could vectorize MaxL/MinL. Probably generally leads to speedups.
- relates to
-
JDK-8351409 Avoid scalar cmov in extreme long min/max loop scenarios
-
- Open
-
-
JDK-8340206 Switch cmov[I, L] x86_64.ad instructs to branch+mov
-
- Open
-
-
-
- Resolved
-
- links to
-
Review(master)
openjdk/jdk/20098