[PATCH RFC bpf-next 0/7] bpf: Fix reg_bounds' slow mode and improve verifier refinement

[Paul Chaignon <paul.chaignon@gmail.com>]

This patchset fixes several failing test cases in the slow-mode
reg_bounds selftests. It also includes an improvement to the verifier
u32->*64 refinement logic, to help better align it with reg_bounds'
refinement logic.

I'm sending as an RFC because it needs [1] to be merged first. The two
patches from [1] are included here, before my fixes.

Patch "bpf: Improve 64bits bounds refinement from u32 bounds" also
happens to address the refinement shortcomings discussed at [2]. To
demonstrate that, I included the new selftests from [2] at the end of
this series.

1: https://lore.kernel.org/all/20260415160728.657270-1-harishankar.vishwanathan@gmail.com/
2: https://lore.kernel.org/all/20260410124035.297632-1-koike@igalia.com/

Harishankar Vishwanathan (1):
  bpf/verifier: Use intersection checks when simulating to detect dead
    branches

Paul Chaignon (5):
  selftests/bpf: Test for empty intersection of tnum and u64
  selftests/bpf: Fix reg_bounds to prune on range violations
  bpf: Improve 64bits bounds refinement from u32 bounds
  bpf: Remove dead code from u32->*64 refinement logic
  selftests/bpf: Hardcode insteresting 32->64 refinement cases

Eduard Zingerman (1):
  selftests/bpf: new cases handled by 32->64 range refinements

 kernel/bpf/verifier.c                         | 299 +++++++++++++++---
 .../selftests/bpf/prog_tests/reg_bounds.c     |  38 ++-
 .../selftests/bpf/progs/verifier_bounds.c     | 202 ++++++++++++
 3 files changed, 482 insertions(+), 57 deletions(-)

-- 
2.43.0