Re: [PATCH bpf-next v2 07/13] bpf: track low-32 scalar equality across sign-extending movs

[Vineet Gupta <vineet.gupta@linux.dev>]

On 9/12/26 12:09 PM, Alexei Starovoitov wrote:
> On Thu Sep 10, 2026 at 9:46 AM PDT, Vineet Gupta wrote:
>> The zero-extending mov records that dst shares src's low 32 bits. A 32-bit
>> sign extension shares them too -- it keeps the low half and fills the high
>> half from bit 31 -- so the same link applies, with a different rule for
>> rebuilding the high bits:
>>
>>    r6 = ...              /* full 64-bit unknown */
>>    r7 = (s32)r6          /* 32-bit sign-extending mov */
>>    if w6 == -1 goto ...  /* taken: r6's low 32 bits are all ones */
>>    ...                   /* r7 is -1, not deduced today */
>>
>> Add SUBREG_SEXT alongside SUBREG_ZEXT, and sext_32_to_64() alongside
>> zext_32_to_64() to drive the reconstruction. Both work from the base's
>> 32-bit range, which is what a 32-bit compare narrows.
>> coerce_reg_to_size_sx() cannot serve here: it reads smin/smax, which
>> straddle after such a compare and collapse to the full field range.
>>
>> tnum_sext() is the counterpart to tnum_cast(). Unlike a tnum_range() over
>> the new bounds it keeps the known low bits.
>>
>> The enum has room for the third value, so bpf_reg_state stays 80 bytes.
>>
>> Unlike the zero-extending arm, a self-mov can form a link here, but only
>> when src is already linked: r0 = (s32)r0 is how a sign-extended int return
>> lands. On an unlinked register there is nothing to link to, and minting an
>> id would leave the register describing itself.
>>
>> Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
>> ---
>> v2: was RFC 5/6.
>>   - no forward declaration (Eduard)
>>   - src renamed known_reg (Eduard)
>>   - sext_32_to_64() and tnum_sext() rather than reusing
>>     coerce_reg_to_size_sx(); the sync path needs the base's 32-bit range,
>>     see the cover letter
>>   - tnum_sext() keeps the known low bits a tnum_range() would drop (Eduard)
>>   - a self-mov links only when src already has an id, narrower than the RFC
>>
>>   include/linux/bpf_verifier.h |  1 +
>>   include/linux/tnum.h         |  3 +++
>>   kernel/bpf/log.c             |  2 ++
>>   kernel/bpf/tnum.c            | 15 ++++++++++++
>>   kernel/bpf/verifier.c        | 47 +++++++++++++++++++++++++++++++++---
>>   5 files changed, 65 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
>> index f1b01059c5da..920c9490ecc8 100644
>> --- a/include/linux/bpf_verifier.h
>> +++ b/include/linux/bpf_verifier.h
>> @@ -53,6 +53,7 @@ enum bpf_add_const {
>>   enum bpf_subreg {
>>   	SUBREG_NONE = 0,
>>   	SUBREG_ZEXT,		/* high bits are zero (32-bit zero-extending mov) */
>> +	SUBREG_SEXT,		/* high bits repeat bit 31 (32-bit sign-extending mov) */
> If my earlier suggestion to model SUBREG_ZEXT as add_const32 with delta == 0
> works, then sign extension could be:
>
> enum bpf_add_const {
>         ADD_CONST_NONE = 0,
>         ADD_CONST_32,    /* delta was added with a 32-bit ALU op */ /* works as ZERO_EXTEND when delta == 0 */
>         ADD_CONST_64,    /* ... with a 64-bit ALU op */
>         SIGN_EXTEND_32, 	/* delta has to be == 0 */
> };

If we were to go that route sure, but no point if we are not doing that.

Also keeping them together makes add_const and sext mutually exclusive. 
Today they are but what if in future we were to implement the following.

w7 = w6      /* r7 gets SUBREG_ZEXT */
w7 += 4      /* r7 also gets ADD_CONST_32, delta 4 */

> and if we really want to we can track movsx 8 and 16 too.
> Which might be an overkill.

Indeed.
Eduard and I discussed this in the off line review before RFC. gcc 
currently doesn't seem to generate these a lot so maybe leave this for 
future.

Thx,
-Vineet