https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116689
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2024-09-12
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed. Note the sink pass only moves to not post-dominated places (and the
sink pass does not consider register pressure for example when sinking stores
or loads).
The sink pass also sinks common code through PHIs to CFG merges (only stores
at the moment), in a separate mechanism.
In principle we could enhance the sink pass to (optionally) sink stmts to
the nearest place where it causes the least register pressure though I
guess doing this greedily bottom-up as sink works does not result in an
optimal schedule for register pressure purposes.
Note on GIMPLE threre's TER which will wreck any scheduling that's done
within TERs bounds (same basic-block, doesn't cross some special stmt
kinds), but scheduling at basic-block granularity might be a win. Since
we already compute liveness at out-of-SSA time we might as well implement
this scheduling at out-of-SSA time.
