https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125305
Mingcong Bai <jeffbai at aosc dot io> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jeffbai at aosc dot io
--- Comment #6 from Mingcong Bai <jeffbai at aosc dot io> ---
For my part (as maintainer of LoongArch port for AOSC OS), I do not think it is
a good idea to forgo instruction scheduling, even at build-time cost.
I can concede that, at -O0, as with AArch64, this scheduling may be disabled -
it seemed reasonable for me - but producing faster and reasonably reliable
binaries at a cost at build-time seems more than reasonable to me. For
instance, AOSC OS builds all of its packages with LTO enabled, this would have
likely been a significantly higher time cost anyway.
(And, just to be sure, I would most definitely NAK the change outlined in
comment 4, as it seems not technically correct...)
I also agree with Xi Ruoyao that, before we go down a more drastic route (or,
as I would argue, even in following AArch64's methodology), we should have
concrete benchmarks to support our decision - whichever it may be.
