https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119010
Bug ID: 119010
Summary: [15 Regression] 444.namd shows a huge compile-time
regression with -mtune=znver5
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
With -Ofast -march=znver5 444.namd compile time (the ComputeNonbondedUtil.C
TU),
sky-rockets from taking 8s with GCC 14 to more than 40s.
A similar slowdown can be observed on other archs, for example icelake:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=904.120.8&plot.1=1086.120.8&plot.2=798.120.8&;
the revision range of that points to early-break vectorization with alignment
peeling being the likely cause.
For the znver5 arch the main time is spent in scheduling 2:
scheduling 2 : 34.26 ( 75%)
the vectorization likely exposes more loops triggering latent compile-time
issues on the RTL side. znver[432] do not exhibit the above issue, nor does
icelake-server.
-march=znver4 -mtune=znver5 is enough to trigger the scheduling issue.
