https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98986
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Version|unknown |11.0
CC| |rguenth at gcc dot gnu.org,
| |segher at gcc dot gnu.org
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Clearly making one order canonical would make sense (subreg, extend, lowpart of
a reg should be singled out from other random exprs of the same kind).
Doesn't help if existing backends do not follow the new canonical order.
On tree/GIMPLE we end up "trying" quite some permutes but on statically
generated decision trees (gimple/generic-match.c) which makes it expensive
on the compiler code-size side but not so expensive on the compile-time side.
(so pre-scanning target .mds and auto-generating parts of combine might
even speed it up ...)