DEV Community

loach
loach

Posted on

PC Engine CD BIOS Devlog: Three Experimental Services in One Boot Path, 18 DEV Entries

I'm developing oss-pce-cd-bios, an open-source PC Engine / TurboGrafx-CD System Card replacement project.

In the November 10 update, I described a local commercial-disc diagnostic reaching BIOS API entry points. That was an observation milestone: the original API implementations were still missing, and no game title screen appeared.

This week's progress moves from observing calls to executing a sequence of three project-defined experimental service bodies on the pinned Geargrafx core.

At the same time, the ordinary development ROM has grown from 15 to 18 implemented entry points out of 81 candidates. None of these numbers represent a verified proprietary System Card ABI.

One experimental image, three real service effects

PR #788 combines three development-only services in a single experimental BIOS image:

Candidate slot Experimental behavior
$02 (CD_BASE prototype) Computes and records project-defined CD address state
$03 (CD_READ link) Consumes that state and performs a guest SCSI READ(6) into work RAM
$33 (EX_DSPOFF) Performs project-defined display-control processing

The standalone experiment executes the chain $02 -> $03 -> $33. On pinned Geargrafx with synthetic CD media, the READ(6) service transfers a real 2,048-byte sector into RAM, where the integration test checks the bytes and the return path. The PR reports cold and warm runs.

The distinction is significant: an earlier trick for exploring boot progress was to replace an unimplemented function with a fake RTS. That can reveal the next call without establishing that the skipped service actually does anything. Here, the test looks for specific changes to service state, RAM content, and video-control state.

But these implementations follow project-owned experimental contracts. They are not established implementations of the genuine CD_BASE, CD_READ, or EX_DSPOFF ABI. The combined experimental image is also separate from the ordinary DEV ROM and from a releasable BIOS.

An A/B/C/D comparison that distinguishes fake progress

PR #790 creates a controlled differential test using matched synthetic disc content, selection metadata, and run budgets across four variants:

Variant Behavior
A: Normal PUBLIC loader Stops at the $02 fallback stub
B: Fake RTS at $02 Advances to the $33 fallback stub
C: Fake RTS at $02 and $33 Advances to the $03 fallback stub
D: Real experimental service bodies Performs three experimental effects, including a READ(6) sector transfer, and reaches the expected clean halt

The comparison payload calls $02 -> $33 -> $03, different from PR #788's standalone chain. Both orderings are described explicitly rather than conflated.

For variant D, the tests observe a 2,048-byte transfer of the selected synthetic sector, service-specific state and shadow values, and the required return behavior. Mutants replacing genuine experimental effects with fake returns must fail the checks.

The PR's published evidence includes cold and warm pinned-core runs. Its own commercial arm was NOT_RUN; it did not quietly search ./rom/ or process private commercial images.

This shows that the experimental code does work under its defined test conditions. It does not show that a real commercial game's calling conventions match those conditions.

Joystick and VDC work in the ordinary DEV ROM

Meanwhile, three project-defined DEV entries have been merged:

  • PR #749: EX_BGON at $2E. Sets the background-enable bit in VDC control register R05. A real-core test exposed a subtle bug in the initial read-modify-write approach: the read port returned stale data. The implementation instead keeps a RAM shadow of the relevant control value.
  • PR #795: EX_JOYSNS at $21. Reads joypad 1 through the HuC6280-visible joypad I/O port $1000 and constructs a project-defined return value from direction/button nibbles.
  • PR #796: EX_SCRSIZ at $23. Writes the low byte of the VDC MWR register (R09) and checks that the tested CPU and memory mapping state is preserved.

Most recently, PR #797 added a same-ROM sequential-call integration test for EX_JOYSNS and EX_SCRSIZ. It runs both call orders, without reseeding key registers between the two calls, so an incomplete state restoration becomes visible at the next call. The PR records eight sequential scenarios, alongside single-service checks. This is regression coverage for the project's DEV contract, not a reference ABI certification.

What about commercial discs?

PR #794 expanded a separate SIMULATED returning-stub experiment to include slot $4A.

In local opt-in traces of two commercial discs, inserting synthetic returns at selected missing entries allowed the observer to see later API entries—$23 on one path and $21 on another. One classifier also flagged a mismatched assumed call prefix, so the result should be read as a trace-level progress signal, not universal classifier success.

Crucially, a fake return does not implement the missing BIOS service. These observations do not mean either disc successfully boots, nor do they authorize private reference-BIOS JSR calls. Commercial media and identifying information remain local.

A wider IPL address range, still not production firmware

PR #745 extends the development IPL loader's LBA calculations into 24-bit intermediate values and 21-bit READ(6) addressing, with synthetic boundary tests. The initial INDEX01 selection remains a separate 16-bit host-derived input. This is not full commercial IPL compatibility.

Other merged work, including PR #769 and PR #771, reinforces fail-closed release checks and the separation between verification-only patches and production inputs. It does not magically turn a debugger guard into runtime safety.

As of November 16, 2026, main snapshot c4a5c048:

Metric Status
Ordinary project-owned DEV entry implementations 18/81
Nonreturning fallback stubs 63/81
Independently verified reference ABI contracts 0/81
Approved reference BIOS JSR entry points 0/81
Owner-approved DEV_CALLABLE entries 0
Demonstrated retail title screens 0
Firmware release eligible false

The next hard problem is not merely making more candidate slots return. It is learning the actual caller-visible contracts, implementing the required behavior without assuming the ABI, and measuring a real game's progress without weakening the safety gates.

The new three-service experiment matters because it puts a measurable difference between “execution reached the next slot” and “the previous service actually performed its intended work.”

Repository: https://github.com/mao2009/oss-pce-cd-bios

Detailed G5 synthetic evidence: https://github.com/mao2009/oss-pce-cd-bios/blob/main/docs/g5-progress-2026-11-13.md

Current ledger and implementation notes: https://github.com/mao2009/oss-pce-cd-bios/blob/main/docs/status.md

This independent project is unaffiliated with NEC, Hudson Soft, Konami, or other rights holders and does not publish proprietary BIOS images, commercial game data, or extracted copyrighted code.

Top comments (0)