From: Alexei Starovoitov <ast@xxxxxxxxxx> Compilers can generate the code r1 = r2 r1 += 0x1 if r2 < 1000 goto ... use knowledge of r2 range in subsequent r1 operations The "undo" pass was introduced in LLVM https://reviews.llvm.org/D121937 to prevent this optimization, but it cannot cover all cases. Instead of fighting middle end optimizer in BPF backend teach the verifier about this pattern. The veristat difference: File Program Insns (A) Insns (B) Insns (DIFF) ---------------------------------- ------------------ --------- --------- ---------------- arena_htab.bpf.o arena_htab_llvm 18656 768 -17888 (-95.88%) arena_htab_asm.bpf.o arena_htab_asm 18523 586 -17937 (-96.84%) iters.bpf.o iter_subprog_iters 1109 981 -128 (-11.54%) verifier_iterating_callbacks.bpf.o cond_break1 110 121 +11 (+10.00%) verifier_iterating_callbacks.bpf.o cond_break2 113 91 -22 (-19.47%) Alexei Starovoitov (4): bpf: Relax tuple len requirement for sk helpers. bpf: Track delta between "linked" registers. bpf: Support can_loop/cond_break on big endian selftests/bpf: Add tests for add_const include/linux/bpf_verifier.h | 12 +- kernel/bpf/log.c | 4 +- kernel/bpf/verifier.c | 90 +++++++++-- net/core/filter.c | 24 +-- .../testing/selftests/bpf/bpf_experimental.h | 28 ++++ .../testing/selftests/bpf/progs/arena_htab.c | 16 +- .../bpf/progs/verifier_iterating_callbacks.c | 150 ++++++++++++++++++ 7 files changed, 298 insertions(+), 26 deletions(-) -- 2.43.0
