https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88398
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hubicka at gcc dot gnu.org
--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the only chance would be for FDO to flag the loop as hot plus assign
it a reasonably low expected niter so GCC could consider "fully" peeling it.
But then whether peeling it that situation is profitable depends highly on
the target (that is, the number of peeled copies that might be profitable).
Considering the various forms of pre-decoded yop caches we have in modern
CPUs peeling might do more harm than good (you only get away with the
mispredicted exit jump, plus you add mispredictions of the peeled exits
in case niter fluctuates a lot).
