Bug fixes: serial-lookup refill collision, SCM tie-off multi-driver, and fetch-merge address mismatch

[liyinrong]

1. Serial lookup serves stale data on a coincident refill write

When a lookup request reaches the data stage in the same cycle a refill writes its line, the single-port data SRAM prioritises the write and the request's read is preempted, so data_rdata holds stale bytes from an earlier read. The output select refill_hit_q && !data_req_q.hit trusted that stale read whenever the tag also hit, so the request was served another line's data.

The tag-hit-plus-coincident-refill case is reachable via the two-cycle lookup latency: a request that missed a line before it was cached can still be in flight to the handler when the earlier refill completes and frees its pending/MSHR entry, so a second, redundant refill is issued for the valid line, colliding with a later request that already tag-hits it.

Fix:

• Serve the in-flight refill data on any refill_hit_q (drop the !data_req_q.hit qualifier on proper_rdata / out_way_o / out_error_o).
• Qualify the refill-payload capture (refill_wdata_q / write_way_q / write_error_q) with the tag handshake (refill_hit_d && tag_valid && tag_ready) so a later request's refill cannot overwrite it while the owning request is stalled in the data stage. out_data_o is already shielded by the data_rsp_q fall-through buffer; this protects the un-buffered way/error outputs.

2. Serial SCM tag config tie-off is a multiple driver

In the gen_scm (latch-based tag SCM) branch, assign sram_cfg_out_tag_o = '0;sat inside the g_ways genvar loop, so the output was driven WAY_COUNT times.

Fix: move the tie-off out of the loop so it is driven exactly once.

3. Unexpected fetch merge under address mismatch

In the MERGE_FETCHES path, prefetch_req_ready[i] OR'd in the arbiter grant prefetch_req_ready_tmp[i], so a fetch port was acknowledged even when a different-address request won the shared lookup mux — that port's prefetch
was consumed without ever being looked up.

Fix: gate the ready strictly on an address match with the accepted lookup request, so a port is acknowledged only when its own address is the one being served.

[Copilot]

Pull request overview

Fixes three correctness issues in the Snitch I-cache datapath/control that could cause dropped fetches, multiple drivers, or stale data being served during coincident refills.

Changes:

• Correct fetch-merge ready signaling so a fetch port is only acknowledged when its own address is the accepted lookup address (MERGE_FETCHES path).
• Fix multiple-driver issue by tying off sram_cfg_out_tag_o exactly once in the SCM tag implementation branch.
• Fix serial-lookup refill collision by always serving refill payload on refill_hit_q (even on tag hits) and by latching refill payload/metadata only on a tag-stage handshake to prevent overwrite during stalls.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/snitch_icache.sv	Fixes MERGE_FETCHES ready gating to prevent acknowledging mismatched-address fetches.
src/snitch_icache_lookup_serial.sv	Fixes SCM tag config multi-driver and corrects serial lookup behavior on coincident refill writes (data/way/error selection + safe payload capture).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.