https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128102
--- Comment #2 from 颜森道 <yansendao at 126 dot com> ---
Bisection result for the needs-bisection keyword.
The regression is introduced by Aldy Hernandez's jump threading rework for
GCC 12 — the series that disabled threading through loop latches until after
loop optimizations — first commit:
01b503871805 "Disable threading through latches until after loop
optimizations." (Aldy Hernandez, 2021-09-09)
followed by the same empty-latch / loop-crossing guards being added to
jt_path_registry::cancel_invalid_paths:
4a960d548b7d (2021-09-23) empty-latch check
d8edfadfc7a9 (2021-10-04) "Path crosses loop header" check
ec0124e0acb5 (2021-10-05) "Path crosses loops" check
All of these land after GCC 11.4 (2021-04), so GCC 12 is the first release
affected.
On GCC 11 the early threader threads the state machine (thread1):
Registering FSM jump thread: (6,10) incoming; (10,11)(11,3)(3,4)(4,5)(5,7)
nocopy; ...
The latch is non-empty (5 insns) and quad_pos becomes constant on the 4
threaded switch paths.
On trunk the early threader rejects the same thread, in order:
FAIL: Thread through latch before loop opts would create non-empty latch
Path crosses loop header but does not exit it
Path rotates loop
and the late threader (after loop opts, where these guards are off) rejects:
FAIL: Did not thread around loop and would copy too many statements
(net_insns * fsm-scale-path-stmts >= max-jump-thread-duplication-stmts,
i.e. 23 * 2 >= 15).
Suggested fix (cost-model only, no loop-form change): the late threader can
thread the loop if the x2 fsm-scale-path-stmts penalty is removed or the
duplication budget raised. Both reproduce GCC 11's codegen and restore the
performance on this machine:
--param=fsm-scale-path-stmts=1 -> 0x30b / 11 cmp / 2.10 cyc/B
--param=max-jump-thread-duplication-stmts=20 -> 0x2f1 / 10 cmp / 2.38 cyc/B
default -> 0xe3 / 1 cmp / 4.12 cyc/B
GCC 11.4.0 -> 0x2c3 / 11 cmp / 2.23 cyc/B
So the likely fix is to exempt multiway-branch (switch) threads from the
fsm-scale-path-stmts penalty in the late pass, or raise the duplication
budget for such threads.
