|
|
|
|
@@ -0,0 +1,502 @@
|
|
|
|
|
# rx65 integrity rematch
|
|
|
|
|
|
|
|
|
|
Branch `agent/w1005/rx65`, starting at `25326f97`. Work is confined to
|
|
|
|
|
`data-d2`. The ten functions are already exact and linked at the baseline.
|
|
|
|
|
Each replacement must remove the assigned carrier, storage overlay, or
|
|
|
|
|
identity helper, preserve behavior, score 100%, and preserve the full DOL
|
|
|
|
|
SHA1 `26116613f624061ba99c8d1a299aaa6efa85670d`. Unsuccessful candidates
|
|
|
|
|
are restored. Configuration and shared headers remain outside this task.
|
|
|
|
|
|
|
|
|
|
Read `f-round.md`, `common.md`, `levers.md` through lever 23, the mwdbg
|
|
|
|
|
README, and the rx56 audit and named trial sources. Scratch sources,
|
|
|
|
|
reports, disassemblies, and trial results are in `build/rx65/`. Each trial
|
|
|
|
|
rebuilds only its owned object, checks its pool, and scores a single-unit
|
|
|
|
|
objdiff report. Allocator captures use this worktree as `--project`.
|
|
|
|
|
|
|
|
|
|
## Baseline
|
|
|
|
|
|
|
|
|
|
Fresh 4.3U build, progress report, and `build/43U/ok` passed. Every assigned
|
|
|
|
|
function scores 100%. The baseline report and objects are saved under
|
|
|
|
|
`build/rx65/` for regression comparisons, including functions with repeated
|
|
|
|
|
`.text` section names.
|
|
|
|
|
|
|
|
|
|
## Trials
|
|
|
|
|
|
|
|
|
|
Results and allocator explanations follow below. A high fuzzy score alone
|
|
|
|
|
does not authorize a retained edit.
|
|
|
|
|
|
|
|
|
|
## Initial clean trials and source review
|
|
|
|
|
|
|
|
|
|
The named rx56 trial diffs were applied to the current sources, rather than
|
|
|
|
|
copying older whole files. All ten reproduced the audit percentages. All
|
|
|
|
|
pools remained identical. Every unit was already linked, and configure.py
|
|
|
|
|
is unchanged.
|
|
|
|
|
|
|
|
|
|
The private mwdbg launcher changes only output confinement: captures stay
|
|
|
|
|
under this worktree, while the compiler, debugger driver, and shared port
|
|
|
|
|
lock are the installed versions. Frozen source snapshots prevent queued
|
|
|
|
|
captures from reading a later trial. Ghidra exports likewise use a private
|
|
|
|
|
project directory under `build/rx65`. The installed shared tools are unchanged.
|
|
|
|
|
|
|
|
|
|
### CDBDatabaseSearchConditionsIsMatch
|
|
|
|
|
|
|
|
|
|
The audit trial has 64/64 instructions and two different immediates, both
|
|
|
|
|
addresses of `type`: source SP+0x18, target SP+0x14. Target `conditions` is
|
|
|
|
|
r30 and `key` is r31; there is no register swap here. The Ghidra output
|
|
|
|
|
agrees on the three scalar outputs and the type buffer.
|
|
|
|
|
|
|
|
|
|
A type-only capacity of four makes the frame too small, 99.875%. Five
|
|
|
|
|
bytes and the full remaining key capacity both reach 100%. Retain the
|
|
|
|
|
full capacity, `CDB_RECORD_KEY_STRING_LEN - offsetof(CDBRecordKey,
|
|
|
|
|
keyStrStruct.type)`, which is nine bytes. A key of at most 31 characters
|
|
|
|
|
can produce eight type characters plus NUL. This also avoids the old
|
|
|
|
|
eight-byte buffer's boundary overrun. The separate maker, game, and date
|
|
|
|
|
outputs retain their target stack locations. 64/64 instructions, diffs 0.
|
|
|
|
|
The existing maker-code pointer cast does not imply a four-byte write:
|
|
|
|
|
`CDBConvKeyStrToMakerCode` copies `CDB_KEYSTR_MAKER_CODE_SIZE` (two)
|
|
|
|
|
bytes with memcpy, so the separate `u16` output has sufficient storage.
|
|
|
|
|
|
|
|
|
|
### CDBConvKeyStrToFullPath_
|
|
|
|
|
|
|
|
|
|
The padding-free audit trial has 173/173 instructions, six stack-address
|
|
|
|
|
differences, and no register differences. Target key/path/location/Wii-ID
|
|
|
|
|
are r25/r26/r27/r28, the data base is r30, and r31 is the zero terminator.
|
|
|
|
|
The target's two epoch buffers start at SP+0x78 and SP+0x88. Existing API
|
|
|
|
|
helper calls preserve the instruction sequence but change those stack
|
|
|
|
|
locations or the frame size. Declaring the first epoch string before the
|
|
|
|
|
file-name buffer gives the correct order with ordinary arrays.
|
|
|
|
|
|
|
|
|
|
The file name needs 13 bytes for eight hexadecimal digits, '.', three
|
|
|
|
|
serial digits and NUL. The old eight-byte array was insufficient. Using
|
|
|
|
|
`sizeof("00000000.000")` and `CDB_KEYSTR_EPOCH_SIZE + 1` for the first
|
|
|
|
|
11-byte epoch string removes the invented prefix and reaches 100%.
|
|
|
|
|
No array exists just to occupy the removed prefix. 173/173, diffs 0.
|
|
|
|
|
|
|
|
|
|
### NCDSetIfConfig
|
|
|
|
|
|
|
|
|
|
The clean union split differs only at instructions 8 and 16: the interrupt
|
|
|
|
|
state is in r29 instead of target r31. The later configuration pointer
|
|
|
|
|
already uses r31. Moving the interrupt local to three declaration positions
|
|
|
|
|
does not change this. A real `RegisterUnofficialVersion` inline routine
|
|
|
|
|
contains the interrupt disable, version registration, flag update and
|
|
|
|
|
interrupt restore. It gives the interrupt local the target r31 and keeps
|
|
|
|
|
all 155 instructions identical. This helper performs an operation and does
|
|
|
|
|
not return an identity value or group unrelated storage.
|
|
|
|
|
|
|
|
|
|
### CDBRecordKeyInit
|
|
|
|
|
|
|
|
|
|
The clean audit trial differs at five array addresses, not registers.
|
|
|
|
|
Target maker/type/game strings start at SP+8, SP+0xc, SP+0x14. Declare game,
|
|
|
|
|
type, maker arrays in reverse order and use the actual string capacities.
|
|
|
|
|
Three bytes hold the maker plus NUL and five hold the game plus NUL.
|
|
|
|
|
`CDBRecordCreateAtOnce` rejects `strlen(fileType) >= 6`, so the type buffer
|
|
|
|
|
must hold five characters plus NUL. A preliminary five-byte type candidate
|
|
|
|
|
also matched but was rejected because it is too short for that caller.
|
|
|
|
|
The retained six-byte type buffer reaches 100%. 65/65, diffs 0.
|
|
|
|
|
|
|
|
|
|
### Address::set_err_msg
|
|
|
|
|
|
|
|
|
|
The audit trial preserves the union's pointer-to-ID default behavior and
|
|
|
|
|
scores 99.875%, with only the final text result in r31 instead of r30.
|
|
|
|
|
Separate final-text locals and an explicitly initialized ID do not fix it.
|
|
|
|
|
|
|
|
|
|
Semantic review: the only caller is `Address::entry_friend`, after
|
|
|
|
|
`FriendListCache::check()` returns NWC24_ERR_NETWORK, or `getLastErr()`
|
|
|
|
|
returns NWC24_ERR_SERVER or NWC24_ERR_FULL. `check()` stores the returned
|
|
|
|
|
error in `mErrCode`. The supported input domain therefore assigns the ID
|
|
|
|
|
in every case. The honest form has separate text and ID variables and
|
|
|
|
|
no default assignment, and scores 100%, 80/80 instructions, diffs 0.
|
|
|
|
|
|
|
|
|
|
Lever 10 proof for the uninitialized ID outside that domain: target
|
|
|
|
|
0x4890 assigns r30 the first text pointer. At 0x490c..0x4938, only error
|
|
|
|
|
-31 writes ID 0x19a and errors -32/-6 write ID 0x1c5. The other paths
|
|
|
|
|
branch to 0x493c without assigning an ID; 0x4940 passes r30 to getMessage.
|
|
|
|
|
This is a target-proven missing assignment, not permission to choose a
|
|
|
|
|
new default. The removed union manufactured a pointer-to-ID conversion
|
|
|
|
|
for that unsupported path. No caller behavior in the supported domain
|
|
|
|
|
changes. Keep this limitation visible during parent review.
|
|
|
|
|
|
|
|
|
|
### Open items after manual trials
|
|
|
|
|
|
|
|
|
|
LetterWriter's caption uses r29 instead of target r30 at two moves.
|
|
|
|
|
Declaration scopes, repeated assignment, a form local, and a friend-cache
|
|
|
|
|
local leave that unchanged. Direct nested calls change evaluation order
|
|
|
|
|
and produce 142/141 instructions, so that form is not retained.
|
|
|
|
|
|
|
|
|
|
The two mailbox helpers need flags in r23 and buffer in r24. Plain locals
|
|
|
|
|
currently get r28 and r23. Declaration order and scopes leave 14 register
|
|
|
|
|
differences. Reloading flags adds two instructions and can change behavior
|
|
|
|
|
if the message aliases mutable work storage, so it is not retained.
|
|
|
|
|
|
|
|
|
|
Board's carrier removal promotes three interrupt states from target stack
|
|
|
|
|
slots +8/+0x10/+0x14 to registers and shrinks the frame by 16 bytes.
|
|
|
|
|
Single-state reuse and branch scopes score 97.678444%, below the audit's
|
|
|
|
|
97.837456%. Increasing type capacity alone leaves that audit score.
|
|
|
|
|
|
|
|
|
|
BrowserWindow's clean constructor omits exactly two target dead stores,
|
|
|
|
|
`stw r0,8(r1)` and `stw r0,0xc(r1)`. There is no register-only explanation
|
|
|
|
|
for those missing instructions. Loop scopes leave 96.969696%; initializer
|
|
|
|
|
and array-clear rewrites are worse. Do not replace the deleted carrier with
|
|
|
|
|
new dummy scalars or a renamed fake object.
|
|
|
|
|
|
|
|
|
|
The MAC target passes SP+0xf as text, clears those same 18 bytes, copies
|
|
|
|
|
six MAC bytes to SP+0x11, then loads the overlapping high and low words
|
|
|
|
|
into r30 and r29. This is actual storage reuse in the target. Separate
|
|
|
|
|
text and u64 output, delayed initialization, explicit shifts, and separate
|
|
|
|
|
text scope do not preserve the target instructions. Best manual clean
|
|
|
|
|
score so far is 93.597404%, from numeric initialization after formatting.
|
|
|
|
|
|
|
|
|
|
## Search and capture state
|
|
|
|
|
|
|
|
|
|
Three 1200-second annealing seeds, 65/165/265, were launched for each open
|
|
|
|
|
LetterWriter, mailbox, Board, BrowserWindow, and MAC function, using frozen
|
|
|
|
|
sources and the unchanged global 24-slot CPU limiter. The private search
|
|
|
|
|
copy fixes definition selection and lets the mailbox search cover both
|
|
|
|
|
existing helpers. No shared search tool was changed. All results still
|
|
|
|
|
require source-level semantic review, exact objdiff, and the full gate.
|
|
|
|
|
|
|
|
|
|
The installed debugger queue was occupied by other workers during the
|
|
|
|
|
manual round. Captures remain queued on the required shared lock. Actual
|
|
|
|
|
priority, degree, masks and coalescing evidence will be recorded below;
|
|
|
|
|
source-shape hypotheses above are not allocator trace measurements.
|
|
|
|
|
|
|
|
|
|
## Compiled trial ledger
|
|
|
|
|
|
|
|
|
|
| Unit | Trial | Function score | Instruction diff |
|
|
|
|
|
| --- | --- | --- | --- |
|
|
|
|
|
| database | clean-audit | 99.96875% | src 0x100 base 0x100 insns 64/64; diffs 2: [48, 51] |
|
|
|
|
|
| conv | clean-audit | 99.96532% | src 0x2b4 base 0x2b4 insns 173/173; diffs 6: [12, 16, 17, 99, 103, 104] |
|
|
|
|
|
| database | type-key-size | 99.875% | src 0x100 base 0x100 insns 64/64; diffs 8: [0, 2, 3, 6, 58, 59, 60, 62] |
|
|
|
|
|
| database | type-extension-size | 100.0% | src 0x100 base 0x100 insns 64/64; diffs 0: [] |
|
|
|
|
|
| database | type-max-key | 100.0% | src 0x100 base 0x100 insns 64/64; diffs 0: [] |
|
|
|
|
|
| ncd | clean-audit | 99.935486% | src 0x26c base 0x26c insns 155/155; diffs 2: [8, 16] |
|
|
|
|
|
| letter | clean-audit | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| key | clean-audit | 99.92308% | src 0x104 base 0x104 insns 65/65; diffs 5: [15, 25, 28, 54, 56] |
|
|
|
|
|
| mb | clean-audit | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| board | clean-audit | 97.837456% | src 0x46c base 0x46c insns 283/283; diffs 49: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|
|
|
|
|
| browser | clean-audit | 96.969696% | src 0x100 base 0x108 insns 64/66; --- insert mine 59:59 base 59:61 |
|
|
|
|
|
| address | clean-audit | 99.875% | src 0x140 base 0x140 insns 80/80; diffs 2: [67, 72] |
|
|
|
|
|
| mac | clean-audit | 88.87013% | src 0x130 base 0x134 insns 76/77; --- delete mine 5:6 base 5:5 |
|
|
|
|
|
| conv | epoch-api | 99.93063% | src 0x2b4 base 0x2b4 insns 173/173; diffs 12: [0, 2, 3, 12, 16, 17, 99, 103, 104, 167, 169, 171] |
|
|
|
|
|
| conv | epoch-key-buffer | 99.93063% | src 0x2b4 base 0x2b4 insns 173/173; diffs 12: [0, 2, 3, 12, 16, 17, 99, 103, 104, 167, 169, 171] |
|
|
|
|
|
| conv | filename-api | 99.96532% | src 0x2b4 base 0x2b4 insns 173/173; diffs 6: [12, 16, 17, 99, 103, 104] |
|
|
|
|
|
| key | key-size-constants | 99.8% | src 0x104 base 0x104 insns 65/65; diffs 13: [0, 2, 3, 12, 15, 25, 28, 54, 55, 56, 59, 61, 63] |
|
|
|
|
|
| key | key-extension-capacity | 99.92308% | src 0x104 base 0x104 insns 65/65; diffs 5: [15, 25, 28, 54, 56] |
|
|
|
|
|
| key | key-full-type-capacity | 99.92308% | src 0x104 base 0x104 insns 65/65; diffs 5: [15, 25, 28, 54, 56] |
|
|
|
|
|
| ncd | enabled-last | 99.935486% | src 0x26c base 0x26c insns 155/155; diffs 2: [8, 16] |
|
|
|
|
|
| ncd | enabled-before-result | 99.935486% | src 0x26c base 0x26c insns 155/155; diffs 2: [8, 16] |
|
|
|
|
|
| ncd | enabled-after-result | 99.935486% | src 0x26c base 0x26c insns 155/155; diffs 2: [8, 16] |
|
|
|
|
|
| key | key-order-sized | 100.0% | src 0x104 base 0x104 insns 65/65; diffs 0: [] |
|
|
|
|
|
| key | key-order-max-type | 99.92308% | src 0x104 base 0x104 insns 65/65; diffs 5: [15, 25, 28, 54, 56] |
|
|
|
|
|
| ncd | version-registration-helper | 100.0% | src 0x26c base 0x26c insns 155/155; diffs 0: [] |
|
|
|
|
|
| letter | caption-direct | 94.893616% | src 0x238 base 0x234 insns 142/141; --- replace mine 42:43 base 42:43 |
|
|
|
|
|
| letter | caption-assigned-twice | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| letter | caption-new-form-local | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| mb | fields-flags-first | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-value-at-use | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| letter | caption-outer-local | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| letter | caption-before-address | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| letter | caption-cache-local | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| browser | browser-loop-scopes | 96.969696% | src 0x100 base 0x108 insns 64/66; --- insert mine 59:59 base 59:61 |
|
|
|
|
|
| browser | browser-index-initializers | 88.30303% | src 0xfc base 0x108 insns 63/66; --- replace mine 27:29 base 27:29 |
|
|
|
|
|
| browser | browser-clear-arrays | 80.742424% | src 0xf8 base 0x108 insns 62/66; --- delete mine 29:34 base 29:29 |
|
|
|
|
|
| mb | fields-buffer-first | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-buffer-scoped | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-flags-reload | 99.540504% | src 0xcf4 base 0xcec insns 829/827; --- replace mine 27:28 base 27:28 |
|
|
|
|
|
| conv | filename-capacity | 99.9711% | src 0x2b4 base 0x2b4 insns 173/173; diffs 5: [12, 16, 17, 114, 149] |
|
|
|
|
|
| key | key-five-char-types | 100.0% | src 0x104 base 0x104 insns 65/65; diffs 0: [] |
|
|
|
|
|
| board | board-one-interrupt-state | 97.678444% | src 0x46c base 0x46c insns 283/283; diffs 58: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|
|
|
|
|
| board | board-interrupt-scope | 97.678444% | src 0x46c base 0x46c insns 283/283; diffs 58: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|
|
|
|
|
| board | board-real-type-capacity | 97.837456% | src 0x46c base 0x46c insns 283/283; diffs 49: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|
|
|
|
|
| address | address-separate-final-text | 99.875% | src 0x140 base 0x140 insns 80/80; diffs 2: [67, 72] |
|
|
|
|
|
| address | address-id-initialized | 99.875% | src 0x140 base 0x140 insns 80/80; diffs 2: [67, 72] |
|
|
|
|
|
| address | address-domain-only-id | 100.0% | src 0x140 base 0x140 insns 80/80; diffs 0: [] |
|
|
|
|
|
| conv | filename-hex-capacity | 99.9711% | src 0x2b4 base 0x2b4 insns 173/173; diffs 5: [12, 16, 17, 114, 149] |
|
|
|
|
|
| conv | epoch-outside-scope | 99.9711% | src 0x2b4 base 0x2b4 insns 173/173; diffs 5: [12, 16, 17, 114, 149] |
|
|
|
|
|
| conv | epoch-before-filename | 100.0% | src 0x2b4 base 0x2b4 insns 173/173; diffs 0: [] |
|
|
|
|
|
| mac | mac-number-after-text | 93.597404% | src 0x130 base 0x134 insns 76/77; --- replace mine 6:7 base 6:7 |
|
|
|
|
|
| mac | mac-number-shifts | 76.41558% | src 0x164 base 0x134 insns 89/77; --- replace mine 10:11 base 10:11 |
|
|
|
|
|
| mac | mac-text-scope | 88.87013% | src 0x130 base 0x134 insns 76/77; --- delete mine 5:6 base 5:5 |
|
|
|
|
|
| conv | filename-sixteen | 99.9711% | src 0x2b4 base 0x2b4 insns 173/173; diffs 5: [12, 16, 17, 114, 149] |
|
|
|
|
|
| conv | epoch-clean-final | 100.0% | src 0x2b4 base 0x2b4 insns 173/173; diffs 0: [] |
|
|
|
|
|
| address | address-caller-domain | 100.0% | src 0x140 base 0x140 insns 80/80; diffs 0: [] |
|
|
|
|
|
| letter | caption-memo-cast | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| letter | caption-cstyle-form | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| letter | caption-reinterpret-name | 99.92908% | src 0x234 base 0x234 insns 141/141; diffs 2: [117, 121] |
|
|
|
|
|
| mb | fields-signed-values | 99.848854% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-unsigned-length | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-return-per-branch | 98.88754% | src 0xcfc base 0xcec insns 831/827; --- replace mine 27:28 base 27:28 |
|
|
|
|
|
|
|
|
|
|
## Reboot recovery
|
|
|
|
|
|
|
|
|
|
The machine rebooted before the pending searches and allocator captures
|
|
|
|
|
finished. The five source candidates, all baseline objects, trial sources,
|
|
|
|
|
Ghidra exports, and trial reports survived under `build/rx65`. No source
|
|
|
|
|
change was left in LetterWriter, mailbox, Board, BrowserWindow or MAC.
|
|
|
|
|
`/tmp` audit scratch was lost, but no further audit replay is needed.
|
|
|
|
|
|
|
|
|
|
Fresh post-reboot object builds reproduced 100% for all ten requested
|
|
|
|
|
functions. Origin advanced on disjoint files; the worker did not merge or
|
|
|
|
|
rebase. A fresh non-quick full gate is being run before committing.
|
|
|
|
|
|
|
|
|
|
Four completed allocator captures survived and independently match whole
|
|
|
|
|
objects emitted by the normal compiler: Database clean/exact and Conv
|
|
|
|
|
clean/exact. The interrupted NCD capture is preserved separately and will
|
|
|
|
|
be repeated. Pending search jobs are restarted with the same seeds and
|
|
|
|
|
sources after rebuilding the target objects.
|
|
|
|
|
|
|
|
|
|
The private debugger transport redirects only this worker's emulator
|
|
|
|
|
listener from port 9001 to 19665 and enables TCP_NODELAY. It keeps the
|
|
|
|
|
installed emulator binary and compiler executable unchanged. The GDB
|
|
|
|
|
script changes only the connection endpoint and output paths. Every
|
|
|
|
|
capture is accepted only after whole-object equality with an independent
|
|
|
|
|
normal-compiler run. No shared debugger files or other worktrees changed.
|
|
|
|
|
|
|
|
|
|
The post-reboot non-quick gate passed for all ten owned units: full build
|
|
|
|
|
passed, DOL SHA1 `26116613f624061ba99c8d1a299aaa6efa85670d`, zero
|
|
|
|
|
regressions, zero new forbidden patterns, zero readability warnings.
|
|
|
|
|
The five clean candidates are checkpointed locally before resuming the
|
|
|
|
|
remaining allocator investigations. This is not a completion claim for
|
|
|
|
|
the remaining five original constructs.
|
|
|
|
|
|
|
|
|
|
## Allocator evidence from completed captures
|
|
|
|
|
|
|
|
|
|
Database clean and exact graphs have identical GPR priorities and colors.
|
|
|
|
|
Virtual r33 is `key`, priority 14 of 15, simplify degree 23 of 24 original
|
|
|
|
|
neighbors; r32 is `conditions`, priority 15, degree 24. Both cannot merge
|
|
|
|
|
with the argument/call registers because of interference. Their first
|
|
|
|
|
available masks are zero, so they request r31 then r30. The buffer fix
|
|
|
|
|
changes stack offsets only. Both whole captured objects equal normal
|
|
|
|
|
project objects for the corresponding frozen sources.
|
|
|
|
|
|
|
|
|
|
Conv clean and exact likewise have the same 42-node color priority.
|
|
|
|
|
`wiiId`/`location`/`fullPath`/`keyString` are virtual r35/r34/r33/r32,
|
|
|
|
|
priorities 39/40/41/42 and simplify degrees 23/24/25/25. Their masks are
|
|
|
|
|
zero before requests for r28/r27/r26/r25. Argument-copy coalescing is
|
|
|
|
|
rejected by interference. The replacement changes stack storage, not
|
|
|
|
|
these already-correct assignments.
|
|
|
|
|
|
|
|
|
|
NCD supplies a direct allocator-guided rematch. In the clean trial, config
|
|
|
|
|
r38 and input r32 are colored first into r31/r30, with simplify degrees
|
|
|
|
|
12/13. The interrupt state is virtual r39, priority 55, degree 17. Its
|
|
|
|
|
available mask is 0xa0000000, r31 or r29; the allocator chooses r29.
|
|
|
|
|
The real registration helper makes that same value virtual r41, priority
|
|
|
|
|
53. It is colored before the configuration-loop temporary r40 claims r29,
|
|
|
|
|
so only r31 is available and selected. Its degree remains 17, and copy
|
|
|
|
|
coalescing with r3 is rejected in both forms. PCode definitions show the
|
|
|
|
|
value comes from OSDisableInterrupts and is consumed by OSRestoreInterrupts.
|
|
|
|
|
Both captures equal the normal compiled objects byte for byte.
|
|
|
|
|
|
|
|
|
|
CDBRecordKey keeps the same input allocation after the array cleanup:
|
|
|
|
|
r32/r33/r34/r36/r37/r38 represent recordKey/epoch/gameCode/serialNumber/
|
|
|
|
|
fileType/recordLocation and end up in r26/r27/r28/r29/r30/r31. Their
|
|
|
|
|
priorities are 26/25/24/22/21/20; simplify degrees are 24/23/15/21/14/12.
|
|
|
|
|
A temporary holding strlen's result first opens r31 at priority 11;
|
|
|
|
|
recordLocation can reuse it. Other saved inputs require fresh saved
|
|
|
|
|
registers and cannot coalesce with their call/argument registers. The
|
|
|
|
|
five different instructions in the clean audit trial are array addresses.
|
|
|
|
|
The final six-byte type candidate and removed redundant cast are both
|
|
|
|
|
captured in `key/capture-exact-final-local` and independently byte-identical.
|
|
|
|
|
|
|
|
|
|
The first C++ capture failed before code generation at compiler address
|
|
|
|
|
0x47e090, with emulator WRITE_UNMAPPED. It reproduces only with the
|
|
|
|
|
`-enc SJIS` option. Diagnostic retries omit that encoding option only;
|
|
|
|
|
the normal comparison compile still uses the original SJIS command, and
|
|
|
|
|
no trace is accepted unless the entire emitted objects are identical.
|
|
|
|
|
The normal project build flags are unchanged.
|
|
|
|
|
|
|
|
|
|
BrowserWindow's honest capture assigns virtual r32 (`this`) to r29,
|
|
|
|
|
priority 35 and simplify degree 26. The inlined member constructor's
|
|
|
|
|
object address r37 takes r30 at priority 30, degree 15; the shared zero
|
|
|
|
|
temporary r40 takes r31 at priority 27, degree 20. All three require new
|
|
|
|
|
saved registers; the `this` and member-address copies cannot coalesce
|
|
|
|
|
with r3 because of interference. These are already the target colors.
|
|
|
|
|
Removing the unused object removes only its two stores; priority changes
|
|
|
|
|
cannot produce those missing stores without another real memory write.
|
|
|
|
|
|
|
|
|
|
The MAC clean trace also has the target saved-register colors. Virtual
|
|
|
|
|
r35/r34 hold the high/low numeric words in r30/r29, priorities 14/15,
|
|
|
|
|
degrees 14/15. The data base r36 takes r31 first, priority 13, degree 14.
|
|
|
|
|
All require new saved registers, and the numeric words cannot coalesce
|
|
|
|
|
with their argument registers because they survive the modulo call.
|
|
|
|
|
The remaining failure is storage and initialization: independent numeric
|
|
|
|
|
storage gets aligned word stores, while the target clears the former
|
|
|
|
|
18-byte text region and later loads words at unaligned offsets +0xf/+0x13.
|
|
|
|
|
|
|
|
|
|
Private debugger support: case-sensitive emulator file lookup required
|
|
|
|
|
symlink aliases under `build/rx65/case-include` for four differently cased
|
|
|
|
|
includes. Original headers and normal build commands are unchanged.
|
|
|
|
|
A page-read cache experiment was rejected because it corrupted trace
|
|
|
|
|
names. An exact-address-and-size read cache, invalidated after every
|
|
|
|
|
compiler continuation and inferior mangler call, was then validated:
|
|
|
|
|
the Database clean object, assigned/simplify/coalescing dumps, and all
|
|
|
|
|
three checked backend dumps are byte-identical to the uncached capture.
|
|
|
|
|
Only that validated cache is used for subsequent long captures.
|
|
|
|
|
|
|
|
|
|
Mailbox clean PCode identifies registration/delay flags as virtual r103/r99
|
|
|
|
|
and their buffers as r73/r72. Flags are colored at priorities 249/252,
|
|
|
|
|
degrees 16, with available mask 0x70000000, so both choose r28. Buffers
|
|
|
|
|
come later at priorities 274/275, degree 18; mask 0x61800000 permits r23
|
|
|
|
|
and they choose it. The flag-to-argument and buffer-to-r3 copies are
|
|
|
|
|
interference-rejected. The allocator has not opened r23 when the flags
|
|
|
|
|
are colored. This explains why local declaration shuffles were ineffective:
|
|
|
|
|
the buffers and flag expressions belong to different groups of inlined
|
|
|
|
|
temporaries.
|
|
|
|
|
|
|
|
|
|
An actual shared formatter added two instructions because its format
|
|
|
|
|
parameter survives the buffer-clear call; rejected at 99.362755%.
|
|
|
|
|
A helper that clears and returns the buffer leaves 99.91536% unchanged.
|
|
|
|
|
Using the declared bitfields moves the delay shift and scores 99.61306%.
|
|
|
|
|
A shared `GetMBFlags(msg, mask)` accessor performs the field load and mask
|
|
|
|
|
inside an inline boundary. It removes the carrier and reaches 100%,
|
|
|
|
|
827/827 instructions, diffs 0, no other function drops. It computes the
|
|
|
|
|
masked field; it does not return an argument unchanged. Its confirming
|
|
|
|
|
allocator capture and full gate are pending.
|
|
|
|
|
|
|
|
|
|
The BrowserWindow loop-scope confirmation has the same complete object,
|
|
|
|
|
priority, simplify and coalescing dumps as the honest baseline: 96.969696%.
|
|
|
|
|
The delayed MAC numeric initialization improves 88.87013% to 93.597404%,
|
|
|
|
|
but its confirming trace retains exactly the high/low/data-base colors,
|
|
|
|
|
priorities and degrees recorded above. Both original forms remain in the
|
|
|
|
|
source tree.
|
|
|
|
|
|
|
|
|
|
LetterWriter clean trace: `this` is virtual r32, first in priority, degree
|
|
|
|
|
12 at simplification, and opens r31. The conversion constant 0x4330 in
|
|
|
|
|
virtual r54 opens r30 at priority 26, degree 14. The picture-branch Board
|
|
|
|
|
pointer r40 opens r29 at priority 40, degree 15. Finally caption r33 is
|
|
|
|
|
priority 47, degree 13, with mask 0x60000000; it chooses the lower free
|
|
|
|
|
register r29. Both caption copies are interference-rejected. The target
|
|
|
|
|
needs r30, so caption must be colored before the Board pointer opens r29.
|
|
|
|
|
|
|
|
|
|
Replace the identity helper with `setLetterHeaderCaption(writer, name)`,
|
|
|
|
|
which calls the virtual caption formatter, gets the input form, and sets
|
|
|
|
|
its caption. This is a complete UI operation with no carrier. Its local
|
|
|
|
|
crosses a real inline boundary and the function reaches 100%, 141/141,
|
|
|
|
|
diffs 0, with zero other function drops. Confirming trace and full gate
|
|
|
|
|
are pending. The six search jobs for LetterWriter and mailbox were stopped
|
|
|
|
|
after these exact manual candidates to release their CPU slots; the three
|
|
|
|
|
seeds for each still-open item remain running or queued.
|
|
|
|
|
|
|
|
|
|
Mailbox confirmation: the accessor moves registration/delay flags to
|
|
|
|
|
virtual r91/r90, priorities 260/261, still degree 16. Virtual r95/r94
|
|
|
|
|
have already opened r24/r23. The flags now see mask 0x71800000 and take
|
|
|
|
|
r23; the buffers at priorities 275/276 consequently see 0x61000000 and
|
|
|
|
|
take r24. Relevant coalescing attempts remain interference-rejected.
|
|
|
|
|
The whole captured object equals both the independent normal object and
|
|
|
|
|
the object from the project build. The seven-candidate quick full-build
|
|
|
|
|
gate passes, with the required DOL hash and no regressions.
|
|
|
|
|
|
|
|
|
|
Board's clean trace explains the memory/register mismatch. `slotInterrupts`
|
|
|
|
|
is virtual r40, priority 81, degree 18, and uses r26 from mask 0x04000000.
|
|
|
|
|
`waitInterrupts` r39 is priority 82, degree 19, and opens r24.
|
|
|
|
|
`recordInterrupts` r38 is priority 83, degree 22, and reuses r24 from mask
|
|
|
|
|
0x07000000. All copies to/from r3 are interference-rejected. None of
|
|
|
|
|
these nodes spills; the allocator still has saved registers available.
|
|
|
|
|
The target instead stores them at SP+0x14/+0x10/+8 and uses `_savegpr_25`.
|
|
|
|
|
Declaration and scope changes cannot by themselves require those memory
|
|
|
|
|
stores. Artificial volatile annotations, address-retention calls and
|
|
|
|
|
one-element carriers are not acceptable ways to force them. The clean
|
|
|
|
|
capture is byte-identical to its saved normal project object.
|
|
|
|
|
|
|
|
|
|
Additional diagnostic correction: the first LetterWriter capture bypassed
|
|
|
|
|
`sjiswrap.exe`, so its later Japanese error literal was UTF-8 rather than
|
|
|
|
|
Shift-JIS. Its requested function agreed, but the whole object did not.
|
|
|
|
|
That capture is provisional only. The corrected diagnostic snapshot
|
|
|
|
|
spells the original literal's Shift-JIS bytes with ordinary octal escapes;
|
|
|
|
|
the comparison runs the actual `wibo sjiswrap.exe mwcceppc.exe` pipeline
|
|
|
|
|
on the original UTF-8 source. Project source and build flags are unchanged.
|
|
|
|
|
|
|
|
|
|
Board follow-up: the repository already has `utility::autoInterruptLock`
|
|
|
|
|
and `nw4r::ut::AutoInterruptLock`. Replacing the three manual regions with
|
|
|
|
|
these real guards preserves restoration on the early return as well as
|
|
|
|
|
normal exit, without inventing a wrapper. The utility guard produces
|
|
|
|
|
283/283 instructions but 97.34276%; its saved state is still promoted to
|
|
|
|
|
r26, and the frame remains 0x90. The SDK guard adds four instructions and
|
|
|
|
|
scores 95.60424%. Both are rejected and the original carrier is restored.
|
|
|
|
|
The plain audit form remains the best honest Board candidate, 97.837456%.
|
|
|
|
|
|
|
|
|
|
The corrected LetterWriter captures now equal the saved normal project
|
|
|
|
|
objects byte for byte. Caption changes from virtual r33, priority 47,
|
|
|
|
|
degree 13, mask 0x60000000 (r29), to virtual r41, priority 39, degree 13,
|
|
|
|
|
mask 0x40000000 (r30). The Board pointer is virtual r39, priority 41,
|
|
|
|
|
and opens r29 afterward. The 0x4330 temporary still opens r30 first at
|
|
|
|
|
priority 26. Caption copies remain interference-rejected. This confirms
|
|
|
|
|
the predicted priority change without adding an instruction.
|
|
|
|
|
|
|
|
|
|
Board search seed 65 found 97.85159%, but its raw result shuffled the
|
|
|
|
|
eight zero stores into an unreadable order. Restoring ascending stores,
|
|
|
|
|
the original comparison direction and `u32` spelling preserves 97.85159%,
|
|
|
|
|
283/283 instructions and 45 differences. Moving all declarations before
|
|
|
|
|
the first assignment preserves that score too. This readable declaration
|
|
|
|
|
order is the new best honest Board trial; it is still rejected and restored.
|
|
|
|
|
|
|
|
|
|
Address confirmation explains the coalescing change. In the audit trial,
|
|
|
|
|
the first message-text value r38 coalesces into messageId r35 because of
|
|
|
|
|
the pointer-to-ID default. Final text r36 is colored first at priority 28,
|
|
|
|
|
degree 12, mask 0x80000000, and gets r31. The coalesced ID then opens r30
|
|
|
|
|
at priority 29. Removing that invalid-domain conversion prevents the
|
|
|
|
|
r38-to-r35 coalesce: first text r38 is now priority 28, degree 13, and
|
|
|
|
|
opens r30 before final text r36 at priority 29. Final text sees mask
|
|
|
|
|
0xc0000000 and chooses r30. The ID remains separate at priority 30,
|
|
|
|
|
degree 13, and can reuse r30. Other argument-copy coalescing remains
|
|
|
|
|
interference-rejected. Both complete captured objects match their saved
|
|
|
|
|
normal project objects, including the accepted exact variant.
|
|
|
|
|
|
|
|
|
|
The real Board utility guard confirmation promotes its three `mOldState`
|
|
|
|
|
values to virtual r117/r116/r115. They are colored at priorities 6/7/8,
|
|
|
|
|
all with simplify degree 17. The first opens r26 and the other two reuse
|
|
|
|
|
it, with no spill and no successful argument-copy coalescing. Its object
|
|
|
|
|
matches the normal project trial byte for byte. Thus even a real existing
|
|
|
|
|
guard changes priority without supplying the target's memory storage.
|
|
|
|
|
|
|
|
|
|
BrowserWindow seeds 65/165/265 each completed their 1200-second budget,
|
|
|
|
|
with 1480/1441/1442 trials respectively. None improved 96.969696%.
|
|
|
|
|
Board seed 65 completed 382 trials, producing the readable 97.85159%
|
|
|
|
|
variant discussed above. The remaining Board and MAC budgets are still
|
|
|
|
|
running or waiting on the unchanged shared 24-slot limiter.
|
|
|
|
|
|
|
|
|
|
Mechanical trial failures were discarded: Mail_sprintf requires a mutable
|
|
|
|
|
format-pointer declaration; the C compiler rejects declarations following
|
|
|
|
|
statements without a new block. A search-result normalization script first
|
|
|
|
|
selected too large a function range and failed compilation; restricting it
|
|
|
|
|
to appendRecord's closing brace produced the measured readable candidates.
|
|
|
|
|
Each trial restored its saved source and rebuilt the original object.
|
|
|
|
|
|
|
|
|
|
The best readable Board capture also matches its normal project object.
|
|
|
|
|
Declaration order makes wait/record/slot state virtual r40/r39/r38,
|
|
|
|
|
priorities 81/82/83 and degrees 19/22/18. Wait opens r24; record and slot
|
|
|
|
|
reuse it from masks 0x07000000 and 0x05000000. The state variables still
|
|
|
|
|
do not spill, so the target's three stack slots remain missing.
|
|
|
|
|
|
|
|
|
|
Board seeds 165 and 265 completed 358 and 318 trials without improving
|
|
|
|
|
97.837456%. All three Board search budgets are complete; the normalized
|
|
|
|
|
seed-65 result remains the best honest score, 97.85159%.
|
|
|
|
|
MAC seeds 65 and 165 completed 1436 and 1386 trials without improving
|
|
|
|
|
93.597404%. Seed 265 remains queued. Its unchanged lock polling was
|
|
|
|
|
resumed after one 0.37-second scheduling delay; the 24-slot limit and
|
|
|
|
|
lock implementation were not changed.
|
|
|
|
|
|
|
|
|
|
## Follow-up compiled trials
|
|
|
|
|
|
|
|
|
|
| Unit | Trial | Function score | Instruction result |
|
|
|
|
|
| --- | --- | ---: | --- |
|
|
|
|
|
| key | key-clean-final | 100.0% | src 0x104 base 0x104 insns 65/65; diffs 0: [] |
|
|
|
|
|
| database | post-reboot | 100.0% | src 0x100 base 0x100 insns 64/64; diffs 0: [] |
|
|
|
|
|
| conv | post-reboot | 100.0% | src 0x2b4 base 0x2b4 insns 173/173; diffs 0: [] |
|
|
|
|
|
| ncd | post-reboot | 100.0% | src 0x26c base 0x26c insns 155/155; diffs 0: [] |
|
|
|
|
|
| key | post-reboot | 100.0% | src 0x104 base 0x104 insns 65/65; diffs 0: [] |
|
|
|
|
|
| letter | post-reboot | 100.0% | src 0x234 base 0x234 insns 141/141; diffs 0: [] |
|
|
|
|
|
| mb | post-reboot | 100.0% | src 0xcec base 0xcec insns 827/827; diffs 0: [] |
|
|
|
|
|
| board | post-reboot | 100.0% | src 0x46c base 0x46c insns 283/283; diffs 0: [] |
|
|
|
|
|
| browser | post-reboot | 100.0% | src 0x108 base 0x108 insns 66/66; diffs 0: [] |
|
|
|
|
|
| address | post-reboot | 100.0% | src 0x140 base 0x140 insns 80/80; diffs 0: [] |
|
|
|
|
|
| mac | post-reboot | 100.0% | src 0x134 base 0x134 insns 77/77; diffs 0: [] |
|
|
|
|
|
| mb | fields-shared-formatter-api | 99.362755% | src 0xcf4 base 0xcec insns 829/827; --- replace mine 27:28 base 27:28 |
|
|
|
|
|
| mb | fields-cleared-buffer-return | 99.91536% | src 0xcec base 0xcec insns 827/827; diffs 14: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 472, 475, 480] |
|
|
|
|
|
| mb | fields-mask-accessor | 100.0% | src 0xcec base 0xcec insns 827/827; diffs 0: [] |
|
|
|
|
|
| mb | fields-declared-bitfields | 99.61306% | src 0xcec base 0xcec insns 827/827; diffs 15: [426, 430, 433, 435, 436, 440, 445, 461, 465, 468, 470, 471, 472, 475, 480] |
|
|
|
|
|
| letter | caption-setup-helper | 100.0% | src 0x234 base 0x234 insns 141/141; diffs 0: [] |
|
|
|
|
|
| board | board-existing-interrupt-guards | 97.34276% | src 0x46c base 0x46c insns 283/283; diffs 69: [0, 2, 3, 7, 12, 19, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35] |
|
|
|
|
|
| board | board-sdk-interrupt-guards | 95.60424% | src 0x47c base 0x46c insns 287/283; --- replace mine 54:55 base 54:55 |
|
|
|
|
|
| board | board-search65-readable-fixed | 97.85159% | src 0x46c base 0x46c insns 283/283; diffs 45: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|
|
|
|
|
| board | board-search65-declarations-fixed | 97.85159% | src 0x46c base 0x46c insns 283/283; diffs 45: [0, 2, 3, 7, 12, 21, 22, 23, 24, 25, 26, 27, 28, 29, 31, 32, 33, 34, 35, 38] |
|