Re: [PATCH bpf-next 1/2] bpf: support shift operations with non-const src operand

[Shung-Hsi Yu <shung-hsi.yu@suse.com>]

Nit: subject prefix is missing 'v3'.

On Tue, Jun 23, 2026 at 12:42:12PM +0000, bot+bpf-ci@kernel.org wrote:
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index 2abc79dbf281..bf247c219cab 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> 
> [ ... ]
> 
> > @@ -14516,14 +14565,14 @@ static bool is_safe_to_compute_dst_reg_range(struct bpf_insn *insn,
> >  	case BPF_MOD:
> >  		return src_is_const;
> >
> > -	/* Shift operators range is only computable if shift dimension operand
> > -	 * is a constant. Shifts greater than 31 or 63 are undefined. This
> > -	 * includes shifts by a negative number.
> > +	/*
> > +	 * Shifts greater than 31 or 63 are implementation-defined behaviour.
> > +	 * This includes shifts by a negative number.
> >  	 */
> >  	case BPF_LSH:
> >  	case BPF_RSH:
> >  	case BPF_ARSH:
> > -		return (src_is_const && reg_umax(src_reg) < insn_bitness);
> > +		return reg_umax(src_reg) < insn_bitness;
> 
> This drops the src_is_const requirement so that non-constant but bounded
> shift amounts are now range-tracked.
> 
> Is there a real use case driving this?
> 
> Earlier review on this series asked whether compilers actually generate
> non-const bounded shifts, and whether any real programs are rejected today
> because of the conservative handling.  It was also noted that the current
> behaviour does not reject such programs, it just falls back to an unbounded
> dst register.  The range tracking added for BPF_DIV/BPF_MOD deliberately
> keeps requiring a constant operand for similar reasons.
> 
> Without a motivating program that is rejected today, is it worth adding this
> complexity to the verifier?

I am with the bot here. This adds complexity in the ARSH code path, and
while LSH/RSH code path does pretty much already support non-const
shift, without a compiler generated program showing this pattern it is
hard to justify moving the patchset forward.