114 lines
11 KiB
Markdown
114 lines
11 KiB
Markdown
# Fixes and engineering changes
|
|
|
|
This inventory describes the delta from upstream commit
|
|
`8b8cf72d1fb4de1d05345de8d5b92ec2d500fcd2`. It distinguishes repairs to inherited
|
|
logic from new features and defects found while integrating those features.
|
|
The checks below establish their stated scopes; they do not turn source
|
|
inspection into a claim of a reproduced field failure.
|
|
|
|
## Inherited secondary joystick-button routing, 2026-11-21
|
|
|
|
The MiSTer top-level binding crossed the two joystick secondary buttons while
|
|
primary buttons and axes stayed with their selected ports. This wiring is
|
|
already present in upstream first commit f00b53c (2021-12-21), not introduced by
|
|
Plus. Correct only the P_SWITCH binding and its misleading secondary-row comments.
|
|
The Tandy 26-3334 service manual, printed page9, assigns PIA0 PA0=right primary,
|
|
PA1=left primary, PA2=right secondary and PA3=left secondary.
|
|
|
|
The extracted actual-RTL regression fails on unchanged wiring and passes524288
|
|
cases on the correction: both port-swap positions, all16 button combinations,
|
|
mouse states, keyboard column masks and keyboard-row patterns. It is included in
|
|
scripts/run_regressions.py. This is a simulation result, not a physical-controller
|
|
or CocoMan firing pass. Native fit, timing and hardware checks are pending here.
|
|
|
|
## Inherited clock, reset and memory paths
|
|
|
|
| Area and problem | Change and relevant source | Verification and limits |
|
|
|---|---|---|
|
|
| FDC/SDC mount metadata used transfer events as clocks, leaving generated-clock and phase hazards. | Capture mount strobes/metadata synchronously, classify on a subsequent phase and present stable consumer inputs. `rtl/fdc.sv`, `rtl/sdc_top.sv`, `rtl/sdc.sv`, top-level bindings. | Interface/phase simulations and fitted setup/hold. A mounted DSK still needs guest-format compatibility checks. |
|
|
| UART timing used independently derived fabric clocks and incompletely coordinated resets. | Use source-clock enables and explicit global/software reset release; preserve UART bit rate and register behavior. `rtl/UART_6551/` and core serial bindings. | Serial/reset component regressions and UART fitted paths. Physical RS232/GPIO electrical routing is separate. |
|
|
| NMI pending/mask state used a generated sampling clock and asynchronous clear path. | Move pending sampling/clear into the master domain, preserve sample ordering and mask/shadow priority, clear pending state on warm reset. `rtl/mc6809glb/mc6809glb.v`. | CPU/NMI regressions and both-corner NMI paths. The S-register change heuristic is retained; this is not a claim of a new fully cycle-exact 6809. |
|
|
| Reset sequencing was timed through a signal treated as a clock, with high-fanout cross-phase release. | Separate reset requests from synchronized consumer releases at the 57/114 MHz and opposite-edge boundaries; preserve original startup counters and peripheral-before-CPU order. `rtl/coco3fpga.sv`, disk wrappers and SDC/WD logic. | Extracted actual reset chains, complete reset timing scopes, negative control for unsupported consecutive enables. No blanket false-path waiver replaces release timing. |
|
|
| CPU-to-SDRAM requests/completions crossed phase boundaries without explicit registered contracts. | Add `rtl/cpu_sdram_bridge.sv`; register request, address/data/control and completion observation in their proper phases. Preserve the proven nonconsecutive CPU-enable contract. | Protocol, actual-controller, boot-control and delayed-data benches; fitted bridge/control/CPU checks. Simulation adaptations preserve logic but do not model physical SDRAM pins. |
|
|
| Keyboard/control and CPU-visible peripheral values could be observed on the wrong consumer phase. | Capture keyboard flags and control/request boundaries, align data-ready observation and SDC CPU buffer access with actual inverted wrapper clocks. Keep the original WD direct read path. | Exhaustive keyboard decode/phase comparison, peripheral and disk-buffer tests with the real wrapper inversion. A passing leaf-only RAM test was insufficient and was replaced. |
|
|
| Video address arithmetic extended into the SDRAM ACTIVE decision, creating an inherited negative setup path. | Snapshot the video address before ACTIVE; retain the existing two-word read burst and wrap/guard semantics. `rtl/sdram/sdram2.sv`. | Original-versus-candidate fetch comparison and corrupted-address negative control, plus fitted snapshot/activation paths. No video-address false path is added. |
|
|
| SDRAM startup/refresh/return behavior lacked an explicit external timing contract. | Match refresh cadence, startup/precharge recovery and return capture to the documented controller/device profile; phase the forwarded clock through the configured DDIO path. | Command scheduler, delayed CPU/video data models and pin-level STA for the conditional Alliance -7 profile. Unknown module speed/PCB delay remains an assumption. |
|
|
| Pixel, audio and board-I2C boundaries were incompletely modeled in STA. | Add source-matched pixel/forwarded clocks and external input/output budgets; constrain both HDMI video clock modes, audio transport and MCP I2C. Preserve bounded I2C protocol timing. `external-io.sdc`, `CoCo3.sdc`, framework audio/I2C bindings. | Actual divider/duty negative control, audio serialization, delayed I2C and 48 substantive external scopes. Optional HPS/muxed/async pins are classified separately, not silently declared timed. |
|
|
|
|
CPU multicycle constraints are restricted to source-qualified register roles and
|
|
proven enables. Unqualified state remains single-cycle. Reset and SDRAM fixes
|
|
retain explicit phase/latency contracts; adding an extra return cycle merely to
|
|
improve fit is not acceptable.
|
|
|
|
## New SSC implementation and integration corrections
|
|
|
|
These are new hardware functions, not fixes to a pre-existing upstream SSC core.
|
|
`rtl/ssc/` implements the host bridge, cartridge glue, TMS7000-family controller,
|
|
AY-3-8913 and SP0256 speech path. `rtl/coco3fpga.sv` supplies CPU decode/audio;
|
|
`CoCo3.sv` supplies menu/runtime ROM input and actual-PLL reset wiring.
|
|
|
|
- **Runtime ROM lifetime:** guest reset originally risked clearing loaded ROM
|
|
validity. Tie cartridge power reset to the actual core PLL-lock release chain,
|
|
preserving valid firmware across guest reset/cold boot. Both original ROMs must
|
|
pass the strict complete-load contract before ready is asserted.
|
|
- **Shared-address ownership:** SSC must replace the inherited video-extension
|
|
register decode when selected. Gate both sides, qualify reads before the
|
|
ordinary bus mux, and disable SSC selection when no capable socket exists.
|
|
- **Command/CPU phase ordering:** latch host transactions and firmware-visible
|
|
transitions on explicit enables; preserve data/status ordering and reset/abort
|
|
recovery. Never interpret idle FF as a missing cartridge.
|
|
- **Original firmware execution:** instruction, immediate-operand, flag, stack,
|
|
timer/interrupt and peripheral semantics follow the actual firmware and chip
|
|
documentation/reference models. In particular FF operand handling is exercised
|
|
by dedicated regressions. The implementation runs the original bytes rather
|
|
than substituting host-side command responses.
|
|
- **Speech and audio completion:** integrate sequencer/ROM/LPC filtering and AY
|
|
tone/noise/envelope output into the legacy audio mix. Busy/readiness must return
|
|
to idle after normal completion or deliberate abort/reset. Mono SSC does not
|
|
replace Orchestra90's stereo route.
|
|
|
|
Original-firmware integration, chip/component fixtures, fitted live-chip path
|
|
audits and Sound Lab listening/transaction checks provide complementary evidence.
|
|
Human listening confirms audibility; it is not an automated waveform-equivalence
|
|
claim for all possible cartridge commands.
|
|
|
|
## Four-slot integration corrections
|
|
|
|
| Defect found during development | Final behavior and check |
|
|
|---|---|
|
|
| HPS decode/RAM allocation on opposing edges left an inadequate transfer path. | Register incoming transfer metadata/data before allocation/writes. Pool tests and fitted MPI paths check the full stage. |
|
|
| Removal/failure overlap could retain allocation or active autoload state. | Removed sockets free blocks; failed/new downloads cannot corrupt neighboring allocations. Concurrent removal/failure regression checks this explicitly. |
|
|
| Duplicate built-in assignment could create ambiguous routing. | Most recently edited socket wins; singleton devices move. Combined Audio can leave its other component in the old socket. Exhaustive router/config fixtures exercise deterministic normalization. |
|
|
| Main sends remembered files before restoring saved layout. | Stage early slot ROMs while Main holds reset, then accept against restored layout. Testing only a default layout missed the Slot4 case; native parser and composed startup fixtures now include non-default layouts. |
|
|
| Normalization acknowledgement could clear Main's initialization reset too early. | Defer publishing migration status until Main's final release. Staging files alone did not fix this second initialization ordering issue. The old behavior is a negative-control failure. |
|
|
| Long labels truncated the usable OSD rows. | Use short labels within the existing OSD width, retain nine stable pages and explicitly show pool limits, assignment requirements and readiness. Native Main menu parser checks every option's status-bit contract and conditional row visibility. |
|
|
|
|
The new router keeps CTS/SCS independent and qualifies read/IRQ/bank behavior
|
|
against the effective socket. Tests extract the actual top/core bindings rather
|
|
than checking only an isolated idealized router. Native menu/startup tests use
|
|
pinned original Main functions; no custom Main binary is required.
|
|
|
|
## Packaging and startup corrections
|
|
|
|
The launcher and embedded display identity are Coco 3 Plus; native MGL `setname` makes settings/media
|
|
use `coco3plus` consistently. Hidden RBF storage keeps the stock core separate.
|
|
Remembered-file records must use the canonical folder and accept the existing
|
|
absolute/relative Main path conventions. Old-prefix replacements must respect
|
|
folder boundaries, not accidentally rewrite a similarly named folder.
|
|
|
|
The installer passes raw structured SCP arguments. Embedding shell quote marks
|
|
inside a modern SFTP path can create literal quotes or fail a spaced launcher
|
|
upload; shell quoting is reserved for the remote shell script itself. It stages
|
|
and checks files before activation, preserves saved settings, and records rollback.
|
|
These are deployment fixes, not defects attributed to upstream FPGA/Main logic.
|
|
|
|
## Corrections that were test issues, not core bugs
|
|
|
|
Idle SSC status FF is expected. Disk BASIC direct-mode `INPUT#` raises ID
|
|
(illegal direct); stored-program reads succeed. A panic attempted after speech
|
|
already completed cannot prove active abort recovery. Mouse JOYSTK queries must
|
|
use the configured port and a progressing BASIC loop. These observations led to
|
|
corrected tests, not speculative RTL edits. An intermediate SDRAM candidate with
|
|
an unnecessary extra READY/return stage produced a green-@ startup failure; it
|
|
was rejected and the first-beat contract restored before qualification.
|