[Bug tree-optimization/122585] [16 Regression] 4-5% slowdown of 433.milc on AMD zen4 since r16-4469-gd6986e06db5eeb

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122585

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
As improvement in general the function-level invariant motion pass invocation
could be made per loop [nest] instead of per function.  Currently a
interchange/unroll-and-jam of an unrelated loop can trigger LIM on another
loop, but it's not going to be trivial to fully restrict LIM in this way.