On Fri, Oct 02, 2026 at 11:39 AM Alexei Starovoitov wrote:
> That's not going to fly.
> After this series there are two implementations of the same thing.
[...]
> Jiong's RFC converted all passes and deleted the old code in the
> last patch:
> https://lore.kernel.org/bpf/[email protected]/
> Pls do the same in one series. One-off patches can be add + commit.
Agreed. v2 will move all passes to the list, turn
bpf_patch_insn_data() into add + commit, and remove
adjust_insn_aux_data(), bpf_adj_branches() and the per-patch adjust
helpers in the last patch. That also covers the bot's comment on
patch 1.
> That's veristat on x86.
> test_progs loads pyperf with BPF_F_TEST_RND_HI32, and s390, riscv,
> powerpc need zext. bpf_opt_subreg_zext_lo32_rnd_hi32() does more
> patches than misc_fixups there, I think, and it has one call site.
Right. On x86 with BPF_F_TEST_RND_HI32:
patches load time
zext misc flag off flag on
pyperf180 4610 930 0.587 s 1.390 s
pyperf600 7620 1530 1.496 s 4.105 s
Patching is 67% of the profile of pyperf180 with the flag, 57% in
bpf_opt_subreg_zext_lo32_rnd_hi32() alone. v2 will report both cases.
> That's unexpected. Pls share the perf report.
perf record -F 4999 --all-kernel -g -- veristat -q pyperf180.bpf.o,
bpf-next 6a75c73eebd4, x86_64 VM:
Children Self
50.66% 0.12% bpf_compute_live_registers
50.43% 0.00% bpf_compute_subprog_arg_access
50.29% 0.67% analyze_subprog
49.47% 8.99% compute_subprog_args.constprop.0
39.08% 8.43% arg_track_join
25.71% 25.67% __arg_track_join
24.97% 0.62% do_check_common
21.49% 0.04% __bpf_patch_insn_data
4.07% 3.99% arg_track_eq
pyperf180 has 2 subprogs, on_event and the global __get_frame_data,
but analyze_subprog() ran 458 times during that veristat run, which
loaded 6 programs counting libbpf's probes.
