https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125855
Roger Sayle <roger at nextmovesoftware dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |roger at nextmovesoftware dot
com
--- Comment #2 from Roger Sayle <roger at nextmovesoftware dot com> ---
Hmm. Perhaps we should just revert my change...
(1) Is the issue with the USE in abs/neg or is this some other USE?
(2) Are the same results observed for AMD and Intel?
(3) Are the rtx_costs of memory accesses accurate on Zen 4?
I'll note that clang-19 avoids referencing the constant pool multiple times.
p.s. with these SPEC regression PRs, might it be possible to also include the
net change for SPECint/SPECcpu around the problematic patch? Without context,
it's impossible to tell if twenty benchmarks got faster, and just one
regressed. It's a bias, as no-one posts/mentions pacthes that result in SPEC
improvements.
Sorry for the inconvenience.
