frontend: Eliminate external_pending launch deadlock (non-external next_id via rd_swacc)

[DanielKellerM]

Problem

Reading next_id launches a DMA. Since #73 the register frontend models next_id/status/done_id as PeakRDL external registers. peakrdl-regblock emits an external_pending latch driving cpuif_req_stall_rd/wr — a whole register-block stall while any external access is outstanding.

#134 then gated the external rd_ack on the decoded req strobe. But external_pending masks cpuif_req, so req is high for only one cycle; once external_pending latches, req drops to 0. If the arbiter grant (arb_ready) is not high in that first cycle, rd_ack = req & arb_ready can never fire again → external_pending is stuck high → the whole reg block is frozen. When the config port shares interconnect ordering with other traffic (e.g. a mem-tile DMA whose config slave shares a demux/NoC path with CPU stores), this frozen block back-pressures and deadlocks unrelated memory traffic.

The pre-0.7.0 (reggen) frontend never had this: next_id was hwext + hwre (a hardware-external register with a .re read strobe) — no external_pending, no whole-block stall; the read stalled only locally via the wrapper's reg-bus ready.

Fix

Model the observation registers as non-external, reproducing the reggen behaviour in PeakRDL:

• next_id: a normal internal register (sw=r; hw=w) with a rd_swacc read-strobe (declared as a UDP) driving the launch — a combinational one-cycle pulse, exactly like the old .re.
• status/done_id: internal (from frontend: Model status/done_id as internal status registers #135).

With no external register, peakrdl emits external_pending = 0 and hardwires cpuif_req_stall = '0 — no whole-block stall, no deadlock, and no assert_bad_ext_rd_ack. The frontend fires the launch off rd_swacc; no ack handshake, no hold latch.

Relation to other PRs

This branch incorporates:

• frontend: Expose the reg frontend as a native APB slave #131 (native-APB frontend) — the APB reg block interface.
• frontend: Model status/done_id as internal status registers #135 (internal status/done_id) — required, since all observation regs must be non-external to fully remove external_pending.

and fixes #134's launch-read deadlock via the next_id → rd_swacc change.

Suggested merge order: #131 → #135 → this. It can also be merged standalone (it contains all three); if #131/#135 land first, this rebases down to just the next_id → rd_swacc change + the rd_swacc UDP declaration.

Test

Validated on a Gwaihir mem-tile DMA (l2_dma L2↔L2 passthrough), where the config port shares the mem-tile demux/NoC path with CVA6 SPM stores. On devel + #134 this hard-deadlocks the SoC; with this fix it runs to completion (SLINK SUCCESS), matching the 0.6.5 baseline, with zero assertions.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[DanielKellerM]

SW C-header note (for downstream consumers)

Since this migrates the reg frontend fully to SystemRDL, downstream SW that consumed the old reggen --cdefines header (idma_reg64_2d.hjson) needs the equivalent from PeakRDL. Two data points:

• cheshire handles this in pulp-platform/cheshire#281: it generates regs/idma.h directly from this RDL with
  peakrdl c-header src/frontend/reg/idma_reg.rdl -b ltoh --type-style hier --rename idma_reg64_2d -P SysAddrWidth=64 -P NumDims=2 -P Log2NumDims=1
  and switches its dma.h to offsetof(idma_reg64_2d_t, …). The moved 2D-dim offsets (0x100) then fall out of the generated struct layout automatically.
• Validated end-to-end on a Gwaihir mem-tile DMA (config port sharing the mem-tile demux/NoC path with CPU stores): with this frontend fix + the generated struct header, l2_dma (1D + 2D phases) runs to SLINK SUCCESS with zero assertions.

Flagging per @-review discussion: it may be worth having iDMA itself ship a c-header make target (e.g. in idma.mk, alongside the existing regblock/raw-header rules) so every consumer doesn't re-derive the peakrdl c-header invocation. Not required for this PR — #281 shows it works from the RDL directly — but it would centralize the naming/param choices.

[DanielKellerM]

Superseded by #153 (frontend/reg-deadlock-fix): a minimal, template-only fix that keeps next_id external and preserves the canonical BLOCKING launch contract, fixing the external_pending deadlock at the rd_ack driver with a pending latch — plus the first directed reg-frontend testbench proving it (non-vacuous: the watchdog fires on the unfixed template). This #147 approach (non-external next_id + rd_swacc) changed the contract to poll-based, could drop a launch under contention, and was untested. Closing in favor of #153.