Shadow blind spots: NMI writes during a compared call, hardware multiplier/divider not compared

Gatchan commented

2026-11-05 06:43:41 +00:00

Owner

Found in #56.

A call interrupted by NMI within the same frame count is compared, but the NMI writes into RAM (e.g. pad_raw $7E:0004) differ between the original and the clone: fuzz of _c0249a alone gave 1 mismatch per version at $0004, a byte the routine never touches. False positives now; possible false negatives if the NMI masks a port difference.
Hardware multiplier/divider registers ($4202-$4206, results $4214-$4217) are not compared: the batch 2 port of UpdatePalAnimPulse leaves the multiplier in a different state than the original (unseen by shadow); DrawTimer (divider) was skipped for this reason.

Scope: compare multiplier/divider state for ports that use it (flag like PORT_WRITES_PPU) and fix UpdatePalAnimPulse if a later read can observe it; handle NMI during a compared call (run both sides with the same NMI, or exclude NMI-written bytes with justification); document in docs/validation.md.

Found in #56. 1. A call interrupted by NMI within the same frame count is compared, but the NMI writes into RAM (e.g. pad_raw $7E:0004) differ between the original and the clone: fuzz of _c0249a alone gave 1 mismatch per version at $0004, a byte the routine never touches. False positives now; possible false negatives if the NMI masks a port difference. 2. Hardware multiplier/divider registers ($4202-$4206, results $4214-$4217) are not compared: the batch 2 port of UpdatePalAnimPulse leaves the multiplier in a different state than the original (unseen by shadow); DrawTimer (divider) was skipped for this reason. Scope: compare multiplier/divider state for ports that use it (flag like PORT_WRITES_PPU) and fix UpdatePalAnimPulse if a later read can observe it; handle NMI during a compared call (run both sides with the same NMI, or exclude NMI-written bytes with justification); document in docs/validation.md.

Gatchan added this to the Phase 2 - field ports milestone

2026-11-05 06:43:41 +00:00

Gatchan added the

bug

validation

labels

2026-11-05 06:43:42 +00:00

Gatchan referenced this issue

2026-11-05 06:53:44 +00:00

Field batch 4: deferred routines of batch 3 and the next field order #59

Gatchan referenced this issue from a pull request that will close it,

2026-11-05 11:00:38 +00:00

Shadow: skip interrupted calls, compare the math unit (fixes #58) #60

Gatchan referenced this issue from a commit

2026-11-05 11:00:41 +00:00

verify: shadow skips calls interrupted by NMI/IRQ, compares the multiplier and divider; UpdatePalAnimPulse bridge leaves the multiplier as the original (refs #58)

Gatchan referenced this issue from a commit

2026-11-05 11:00:41 +00:00

verify: UpdateShake bridge leaves the multiplier, ClearTextGfxBuf bridge the WRAM port, as the original (found by the #58 shadow) (refs #58)

Gatchan closed this issue

2026-11-05 11:00:55 +00:00

Gatchan commented

2026-11-05 11:00:56 +00:00

Author

Owner

Merged in PR #60. Chapter 1 routes, 46 ports in shadow: US 1 552 349 calls 0 mismatch 0 fallback (9 567 skipped), JP 1 407 933 calls 0 mismatch 0 fallback (7 621 skipped); fuzz _c0249a alone 0 mismatch both versions. The new math-unit comparison found two more bridges to fix (UpdateShake multiplier, ClearTextGfxBuf WRAM port), fixed in the PR.

Gatchan referenced this issue from a commit

2026-11-05 11:00:59 +00:00

Merge pull request 'Shadow: skip interrupted calls, compare the math unit (fixes #58)' (#60) from shadow-nmi-math into main

Rows
Columns

Shadow blind spots: NMI writes during a compared call, hardware multiplier/divider not compared #58