Re: [PATCH v2 bpf-next 1/2] bpf: Avoid one round of bounds deduction

[Eduard Zingerman <eddyz87@gmail.com>]

On Mon, 2026-03-09 at 18:07 -0700, Eduard Zingerman wrote:

[...]

> Of-course, I might have some bugs in the test definition, but in case
> the definition is correct, I think I'm inclined to agree with Ihor
> here. If we can't mathematically prove that the sequence converges in
> a fixed number of steps, putting a loop on top of it seem to be a
> reasonable thing to do.
> 
> [1] https://github.com/eddyz87/bpf/tree/deduce-bounds-reshuffle
> [2] https://github.com/eddyz87/deduce-bounds-verif
> 
> [...]

Pushed an update to the test to check if new bounds are tighter than
old bounds (instead of equivalent). Still get a failure log:

$ ./parse-log.sh out2.log
Values extracted from log:
  s32: [-1686110209,  1907957760]  v_s32=-1411416065
  u32: [ 1746894847,  4030210048]  v_u32=2883551231
  s64: [ -4350478693995380738,   2172132303447261182]  v_s64=-4350478693150261249
  u64: [  2172132302103053322,  14096265381705949184]  v_u64=14096265380559290367

Running C code directly:
input:
  s32: [0x9b7fffff, 0x71b92000] [-1686110209, 1907957760]
  u32: [0x681f7fff, 0xf0382000] [1746894847, 4030210048]
  s64: [0xc39ffead797ffffe, 0x1e24f6d2501ffffe] [-4350478693995380738, 2172132303447261182]
  u64: [0x1e24f6d20001040a, 0xc39ffeadf0382000] [2172132302103053322, 14096265381705949184]

old:
  s32: [0x9b7fffff, 0xf0382000] [-1686110209, -264757248]
  u32: [0x9b7fffff, 0xf0382000] [2608857087, 4030210048]
  s64: [0xc39ffead9b7fffff, 0xc39ffeadf0382000] [-4350478693424955393, -4350478692003602432]
  u64: [0xc39ffead9b7fffff, 0xc39ffeadf0382000] [14096265380284596223, 14096265381705949184]

new:
  s32: [0x9b7fffff, 0x71b92000] [-1686110209, 1907957760]
  u32: [0x797ffffe, 0xf0382000] [2038431742, 4030210048]
  s64: [0xc39ffead797ffffe, 0xc39ffeadf0382000] [-4350478693995380738, -4350478692003602432]
  u64: [0xc39ffead797ffffe, 0xc39ffeadf0382000] [14096265379714170878, 14096265381705949184]