Repository navigation
Description
A narrow read through the 64-bit AXI / 32-bit register bridge can also read an inactive adjacent register. If that register has a read side effect, the extra access changes device state. The bridge currently filters zero-strobe writes only.
The candidate also acknowledges zero-strobe read banks locally. It requires the companion AXI read-byte-enable correction and must not be used with the old read-strobe behavior. The old converter takes read strobes from the unrelated W channel, so the bridge filter alone can suppress valid reads. The Bender dependency stays unchanged. A released compatible AXI version or an explicit compatibility contract is still needed before a ready-to-merge bridge PR.
Candidate branch: https://github.com/lionelZhaowy/register_interface/tree/6a5afb3652110344c0ec8d89ef4c00b3c8e12c29
The candidate, public test and exact commands are in test/NARROW_READ_BANKS.md. Companion AXI PR: pulp-platform/axi#463 (candidate 97608408d0f6cc578f4c661fedba83205ce4df36). This report adds functional side-effect evidence; it is distinct from #56/#66's throughput/buffering work.
Across the same 352 cases, the original bridge reports 588 errors. The AXI fix alone leaves 140 side-effect errors. Combining the AXI fix with the bridge filter gives zero errors.
Validation, fixed inputs and scope
Actual official inputs: register_interface master d6e1d4cdaab7870f4faf3f88a1c788eaf5ac129d, AXI master d436bd4936ba1c49cd87521df54580c5159114d3, AXI's declared common_cells db42769334b4589b4b3fc671b34513bdb98be565, and common_cells 1.39.0 9ca8a7655f741e7dd5736669a20a301325194c28 for the unchanged legacy bridge cells. No dependency is upgraded.
Same 352-case test through the real complete bridge:
- Original: 588 errors.
- AXI byte-enable change alone: 140 errors remain from inactive MMIO side effects.
- AXI plus bridge filter: zero errors.
- 96 reads cover size 1/2/4/8 bytes, offset 0..7, INCR lengths 1/2/4 and R backpressure. An independent address-interval oracle checks all downstream register accesses. All 256 write strobes, payloads, response IDs/data/LAST are checked.
- All runs reach normal finish with finite HDL/wall limits and assertions enabled. The side-effect-counting peer is behind the actual bridge; it does not replace the DUT.
- Full-bridge generic DC
compile -no_map: 1095/1087 flip-flops, 11653/12118 cells, zero latches, combinational loops, multiple drivers, unexplained cells or final blackboxes before/after. This is a structure check, not technology timing or area signoff.
The companion direct AXI test also reproduces 448 byte-enable errors before and zero after, independent of this bridge. All candidate material is public and retains the original source license and authors.
OpenAI Codex assisted with implementation, test and validation. A separate AI review checked the changes and results.
Activity
[-]Narrow AXI reads reach inactive register banks[/-][+]Narrow AXI reads can access inactive register banks[/+]