psxrecomp/docs/internal/WO6_BIOS_INTEGRATION_REVIEW.md

WO-6 / WO-7 and BIOS #343 / #346 integration

Tracked by beads-eio.3.140. Integration branch: integrate/wo6-bios343-346. Contributor heads c2f9c3d5 (#343) and baca0a8a (#346) are retained as ancestors, not copied or squashed. Initial baseline: 736c002a; refreshed baseline: b4ea4c37.

Changes and conclusions

• WO-6: generate an unconditional immutable physical-word index for resident dispatch tables spanning less than 2 MiB. Wide/ambiguous tables retain the original binary search. Keep CPS returns, live-code validation, interrupt checks and mod hooks. This accelerates address resolution; it does not turn overlay calls into nested C calls or promise fewer dispatcher activations. Tomba's 18,130-entry index is 265,706 bytes; MMX6's 18,858-entry index is 255,918 bytes. An isolated benchmark of the actual emitted lookup with simplified table records and 4,096 repeated mixed-hit queries measured roughly 54-61 ns/query for binary search and 1.2-1.6 ns/query for the index. This is not a measurement of total dispatch overhead or game speedup.
• WO-7: the native hot-owner sampler counts owner activations, including CPS continuation returns. Function-entry tracing counts invocations. Its direct-mapped bucket resets on a collision, so it can undercount but does not accumulate another owner's calls. Clarify the telemetry label and API docs; no counter algorithm change or confirmed counter corruption. The supplied brief's 3.2 microseconds divides total guest work by activations and therefore does not isolate dispatch cost. Its branch at 0x80191BC0 with immediate FFFB targets 0x80191BB0, not 0x80191BB4.
• #343: setup detection now requires a configured dispatch/full pair and its matching backend descriptor, and uses the configured framework location. Preserve expected-retail discovery for OpenBIOS;SCPH1001 and alternate retail stems. Add the CLI test file registered but missing from the PR.
• #346: guard patched BIOS ranges before instruction charges and terminators, including entry through interior labels. Compare the whole declared range and hand off at the requested RAM PC. Reject native emission when the range begins across a live branch/load-delay boundary. Do not hand back from the interpreter with a pending load. Clear range metadata on memory reset. No patch ranges were widened and no guest patch effects were synthesized.

Validation

Windows native Clang 22, Release recompiler / RelWithDebInfo runtimes; fresh game and BIOS generation, isolated builds, overlay caches and memory cards. Real title-owned mods were built/staged; no original game output or save was modified. Live runs use LLE BIOS boot/calls, headless framebuffer readback, debug diagnostics and native GCC/Clang overlay compilation. UI, Vulkan, netplay, rewind and PGXP are disabled in this regression configuration.

Recompiler CTest: 65 enabled tests pass, including the executable generated lookup check (aliases, holes, alignment, bounds, stale-code rejection and interrupt/call behavior). Runtime CTest: 78 executed tests pass; one explicit skip and two disabled. Recompiler has three disabled tests; its disabled overlay-pair test was run directly and passed all five executable scenarios. The #343 host probe was also run with real recomp-ui headers and native CC: seven positive/negative cases pass. Upstream relocation/AOT tests: 4 + 24 pass.

The test-harness repairs include two Windows text/byte fixtures, a stale assertion that demanded unsafe continuation-range clipping, a cache-tag fixture now using the canonical serializer, and a Windows stack allowance for the mod-runtime test's real 1 MiB disc-hashing scratch buffer. None changes production guest behavior.

Both linked BIOS images were regenerated. OpenBIOS emits 651 functions with zero tagged interpreter fallbacks and four existing unsupported-instruction skips; retail emits 1,309 with its one existing load-delay fallback and one existing unsupported-instruction skip. Retail SHA-256: 71af94d1e47a68c11e8fdb9f8368040601514a42a5a399cda48c7d3bff1e99d3.

Initial matrix: all eight combinations of Tomba/MMX6, OpenBIOS/SCPH-1001 and baseline/integration completed at least 11,000 frames. Screenshots show real FMV and MMX6's gameplay demo; integration has native overlay/kernel execution, zero remaining kernel-body mismatches, and completed 128-byte card reads. The first baseline Tomba screenshot exceeded a five-second TCP deadline; its rerun passed. The first integration MMX6 run spent most of its budget in host startup and reached only 1,414 frames; a longer-budget retry passed. The baseline also exhibited slow host startup. Do not present either failed attempt as a successful run or interpret cold-start times as dispatch cost.

The rebuilt eight-way matrix against b4ea4c37 passed, with clean process exits and 32 completed 128-byte card transactions per run:

Game	BIOS	Baseline frames	Integration frames	Kernel mismatches, baseline -> integration
Tomba	OpenBIOS	11,005	11,015	22 -> 0
Tomba	SCPH-1001	11,009	11,008	11 -> 0
MMX6	OpenBIOS	11,002	11,013	7 -> 0
MMX6	SCPH-1001	11,014	11,015	11 -> 0

Each integration run has native overlays plus native kernel execution and 11-33 patch-skip events. Observed warm wall times were 82.5/86.2 seconds for Tomba/OpenBIOS, 71.6/74.2 for Tomba/retail, 70.2/69.3 for MMX6/OpenBIOS and 70.6/70.8 for MMX6/retail (integration/baseline). These runs share a busy host and are not controlled FPS benchmarks. Retail Tomba's frame-1,500 checkpoint falls in a display-disabled boot interval on both builds; subsequent checkpoints show the same FMV. Local raw reports/screenshots are under F:/Projects/psxrecomp/_validation-wo6-bios/{baseline,integration}/{tomba,mmx6}/results/; final-openbios and final-SCPH1001 identify the refreshed runs.

An additional OpenBIOS/MMX6 run on each side captured the demo fade-out at roughly 20-frame intervals through frame 11,150. Both show the same fade, brief display-disable interval, and return to the 512-wide Capcom screen. This explains the darker initial integration screenshot sampled 11 frames later than baseline; it is not missing background rendering.

Integration PR: #348.

Scope of confidence

This is regression evidence for the tested boot/FMVs/attract paths, generated dispatch contracts, BIOS patch boundaries and card protocol, not a full playthrough, hardware differential proof, audible-audio assessment, UI test, save/load round trip, cross-platform certification or Parasite Eve FPS result. Tomba 2 and alternate retail BIOS images were not part of this matrix.