Worker F's 0x8002622C (44 B) — A RECORDED NEGATIVE OVERTURNED, with a new class.
The old record said 'cc1 folds it, unreachable' and tried FOUR ALGEBRAIC re-spellings. All four
were doomed: the fold is at RTL combine, not in the front end, so no re-spelling can avoid it.
Only LIVENESS can. Same body with 'return 0' is 32 B LENGTH-MISMATCH; with 'return n' (the
difference live past the addition) it is 44/0/MATCH. Diagnostic that proves the pass: cc1 -da
shows the minus present in the .flow dump and gone in the .combine dump, while cse/cse2/jump/
loop/sched/sched2 all still contain it.
181: when an original keeps an arithmetically-cancelling pair (subu+addu, x-c+c), the intermediate
is LIVE PAST the second operation — find the later reader.
182: a negative with a NAMED mechanism is overturnable; one without is not. 'cc1 folds it' is not
a classification; 'RTL combine cancels it, and algebraic re-spelling cannot reach combine' is,
and it immediately implies the liveness lever.
179: a rule every worker must follow belongs in a TRACKED tool, not a copied script. The registry
free-check lived in a worker's staging dir and the orchestrator told everyone to copy it. It was
WRONG -- it reported TAKEN if ANY ledger row for the address was not 'released', so a row that was
wip and later released stayed blocked FOREVER. Worker E found it and measured 13 released rows
reading as taken, several of them the cheapest rows left, and the same stale pattern existed for
workers A and D too. Fixed and promoted to tools/sf3_free (tracked, docstring explains the bug).
A copied script cannot be fixed for the people who already copied it.
180: worker E's maspsx=epilogue token table, MEASURED not inferred -- 5 rows REQUIRED, 3 HARMFUL
or NO-OP, 2 HARMFUL-but-unmatched. Three of E's fourteen claims sit on that list and TWO would
have failed outright if the token had been applied by shape. The cheap read: '4 bytes SHORT with
the token on' means the token was unnecessary.
609 regions / 600 distinct bodies, from the 484 / 493 baseline at phase start (+116 regions,
+116 bodies). Worker E's 0x80077D04 (120 B) is the 600th.
VERIFIED, not asserted:
full-binary gate : 609 regions, differing_bytes=0, result=MATCH, exit 0
make check : 253 tests, OK, exit 0
every source file : present on disk and tracked by git
working tree : clean
Corpus maximum 1232 B; 28 regions exceed the 244 B ceiling the phase was planned around.
Delivered by six workers over the phase: A 46, D 21, B 17, C 13, E 14, F 3 = 114 claims,
every one gated before merge.
The phase's premise was falsified early and replaced with working tooling:
- the 244 B ceiling was a DISPATCH ARTEFACT (5 of 427 rows above it had ever been attempted);
- ASPSX does not fill delay slots at all, so the authorised post-pass shrank from a modelling
project to one mnemonic (maspsx=moves) plus an epilogue swap (maspsx=epilogue);
- cost is TIE-BREAK DENSITY, not size, measured independently by two workers from opposite
directions;
- the GTE class moved from BLOCKED to OPEN (worker D's 0x800F3E18 is the first GTE row matched
in this project).
175: mask chains must be SEPARATE STATEMENTS -- one expression folds to a single and; three
statements emit the original's lw/and/and/and/sw. cc1 does not fold constants across statements.
176: a MECHANISM, not a tip. Three separate RMW statements give one lw, three ands and ONE sw
(cse forwards each load from the previous store within a block); an intervening early return makes
that store land in the beq delay slot and the tail's store stays alive because its load is
forwarded from a store in the PREVIOUS block. A local chain loses the delay-slot store (100 B);
volatile keeps both stores but flips the entry branch (100 B).
177: fold REASSOCIATES | -- one statement becomes v | (C1 | X), a 17-byte residual; two statements
give the original (v | C1) | X.
178: finding 170 has a CONSTANT-clobber variant -- the same entry-block copies arise because mask
constants are materialised into the parameters' own homes, with no call involved. The rule
generalises: a redundant entry copy means the value is live when its home is clobbered, whether by
a call's argument setup OR by constant materialisation.
172: a MISSING sign bias is evidence of a shift, not a division -- the values are non-negative in
the source but cc1 cannot know that after a conditional, so the absence of the addiu is the
diagnostic. Finding 97 read backwards.
173: the absolute value is a SWAPPED SUBTRACTION (subu with operands exchanged), not negu -- the
natural -x is the trap. Second instance of finding 101.
174: the idiom-redundant-by-construction class, and the discipline of declining a row. Worker E
decoded 0x800FECF8 -- one of the highest-scoring fresh small rows -- and RELEASED it without
attempting it, because it is a magic-division row where every division has several equally
plausible spellings. A high redundancy score is not sufficient if the row is in a class known to
be spelling-ambiguous.
Worker E's 0x8005E17C and 0x8002FAB8; worker F's first two claims 0x800FBE84 (216 B, FIRST
SPELLING with worker A's derivation) and 0x80026274 (108 B).
170 generalises the argument-evidence levers (157/164) into a mechanism: a redundant ENTRY-BLOCK
copy of an argument means that value is still live at a call whose argument setup CLOBBERS that
same register. The copy is materialised in the entry block because the tie to a0's home is
illegal. The test that nailed it: the same body with a 2-arg call is 104 B LENGTH-MISMATCH; with
the 3-arg call it is 108/0. Two prior corpus instances had the copy AT the call; this is the
hoisted-to-entry variant.
171: worker F confirmed EXHAUSTIVELY that the constant-division divisor is unique per magic --
(n*M)>>(32+s) == n/D has exactly one D. So finding 67's identity is not an approximation.
Worker A calibrated sf3_family by checking two 0.97 entries and finding neither shared its
sibling's body. I have now confirmed that on eight candidates by raw-word diff, which is the
decisive test: 0x8006EBA0 vs 0x80028CE0 differs in 60 of 61 words; 0x800FFFEC vs 0x8007E8B8 in
18 of 19; 0x8005E17C vs 0x800FB54C in 25 of 26. Contrast the genuine sibling 0x800F3DC0 vs
0x800F3E18: 1 of 22 words.
So a high cosine with ratio 1.00 is NOT evidence of a shared body -- at 0.96-0.99 the histogram
matches common IDIOMS. The useful band is ratio 1.000 AND a near-zero raw-word diff.
Tool bug fixed: sf3_family did not exclude already-claimed rows, so its top hit was a row
matching ITSELF. The registry is now the authority and claimed rows are skipped.
Worker E's 0x8002311C (160 B) CLOSES COOKBOOK 123'S OPEN QUESTION. Finding 123 recorded the
branchless MAX0 (x & -(x > 0)) as unreached -- 'no ternary and no bitwise spelling reached it'.
Worker E solved it: the lever is NAMING THE BOOLEAN.
return s & -(s > 0); -> BRANCHES
return s > 0 ? s : 0; -> branches
flag = s > 0; return s & -flag; -> EXACT (slt / negu / and)
Mechanism: naming the comparison forces cc1 to materialise it as a VALUE (slt) rather than a
test feeding a branch. That is finding 44's 'name the boolean' lever applied to the MAX half --
finding 44 previously had only the cond-into-&& direction for this family.
167: a 4-byte store cc1 DELETES means the object's address is never taken -- fold the word into
the array whose address IS taken by a call.
168: s = f(); s += f(); s += f(); loses one instruction vs three named results summed.
0x800F3DC0 (88 B) — a ONE-WORD sibling of the matched 0x800F3E18, found by worker E via
sf3_family at ratio 1.000 and confirmed by raw-word diff: identical in all 22 words except the
COP2 command field (0x4B70000C vs 0x4B78000C). The route was one copy, two renames and one field
change; every __asm__ and register binding carried over untouched.
166 records it, and notes it is the MIRROR of finding 161: on 0x800F3E18 the field 0x178000c was
the WRONG answer (one byte off, 0x170000c correct); on 0x800F3DC0 0x178000c IS correct. A count
tells you a field is COMMON, not that it is right -- and a ratio-1.000 sibling is the cheapest
place to learn which one a row wants. When the family tool reports one, diff the raw words FIRST.
Records worker output (104 claims total), the GTE class moving from BLOCKED to OPEN (worker D's
0x800F3E18 is the first GTE row matched in the project), the epilogue post-pass and its two
corrections, seven defects in coordinator-written rules, the central finding restated with worker
D's seven-finder breakdown, and the ranker's known blind spot with both failed proxies.
Worker E's 0x80107DE8 (128 B, maspsx=epilogue) and 0x8007D5FC (132 B, DEFAULT toolchain).
163: a recurring epilogue-list class WITH A SHAPE TELL -- correct length, identical instruction
multiset, and the residual is where cc1's reorg put the EPILOGUE LOADS relative to the last
gp-relative read-modify-write block. The original's epilogue loads FILL the global load's delay
slot; cc1 emits a #nop instead. Tell: the row's last statement is a gp-relative read-modify-write
immediately before the epilogue. NOT a spelling problem -- worker E probed the maspsx mode split
and the nop is a cc1 #nop, not maspsx, so no option changes it. Classify it: post-pass territory.
164: a register residual can be ARGUMENT EVIDENCE -- when the only residual is a value in an
argument register and there is no argument setup at the jal, try passing it as that argument.
165: finding 147 confirmed a second time -- 0x8007D5FC is on the epilogue list but cc1 fills its
own slot, so the token is a no-op and the row is claimed with '-'. The list was selected on the
ORIGINAL's tail; the token must be decided from the CANDIDATE's.
Worker A's final row 0x8010A6C4 (132 B, first attempt, maspsx=epilogue) -- its ninth epilogue
row and its 46th claim.
162: a callee called with DIFFERENT argument counts needs a NON-PROTOTYPE declaration --
func_8010A444(1) / (2, x) / (3, s1, s0) is only expressible as 'void func_8010A444();', the C89
empty-parameter form, not '(void)'. Same constraint that cost worker A a compile on 0x8002DD14.
Worker A's final totals: 46 claims (33 first-attempt), 95 evidence rows, 39 levers, 3 deferred
rows with derivations, 1 blocked row, 9 rows carrying maspsx=epilogue.
Worker D's 0x800F3E18 (88 B) -- THE FIRST GTE/COP2 ROW MATCHED IN THIS PROJECT.
159: lwc2/swc2 move a word straight between MEMORY and COP2, unlike mtc2/mfc2 which move
between a GPR and COP2. The row uses lwc2 $9/$10/$11 and swc2 $25/$26/$27, so a row can use
the IR/MAC registers WITHOUT the IR/MAC macros. Added gte_lwc2IR1/2/3 and gte_swc2MAC1/2/3.
160: register variables PIN the COP2 operand registers -- worker D's entire residual was that
cc1 chose its own cfc2/mfc2 destinations. The GTE analogue of the named-locals family: an
inline-asm row's residual is usually the operand REGISTERS, not the sequence.
161 IS A CORRECTION TO THE COORDINATOR'S OWN ADVICE. I broadcast the command-field values with
their occurrence counts as if they were a lookup table. They are a DISTRIBUTION, not a per-row
answer: worker D wrote 0x178000c (counted 51x) into the row and it came out ONE BYTE wrong; the
correct field is 0x170000c. The low bits carry the shift/matrix/vector selectors, so two commands
differing only there are different instructions. Read the field off the ORIGINAL WORD.
Worker A's 0x800F4B88 (128 B, first attempt) -- its eighth epilogue-class row and its last.
158: two type views over the same halfwords are DELIBERATE. The first helper call loads with lh
(signed) and the second with lhu (unsigned) over the SAME pointer, so the source declared a
short* view for one expression and an unsigned short* view for the other. Writing the whole row
as short* gives lh for the second call too and changes the bytes. When one function reads the
same field both ways, the mixed lh/lhu pair over one pointer is the evidence.
Worker A's three epilogue rows (0x800F452C, 0x800F6DD0, 0x800F6E50).
155 is a DISPATCH finding: the epilogue list is ALSO a family list. 0x800F6DD0 and 0x800F6E50 are
siblings differing in exactly two ways, and worker A read one and got the second for free, both
first try. Adjacent pairs already identified: 0x800F6DD0/0x800F6E50, 0x800F42AC/0x800F452C,
0x800FFFEC/0x80100038. A worker taking an epilogue row should read its NEIGHBOURS first -- the
class was selected on a TAIL SHAPE, and tail shape correlates with the translation-unit layout
that makes neighbours siblings. Generalised: any class selected by a structural feature clusters
its results by address.
156: the three writes are ASSIGNMENTS not accumulations -- the original never loads the old
destination value, so writing += adds three loads.
157: fewer argument registers set than parameters means the source passes its OWN LIVE parameters
directly. Now confirmed on three rows.
Worker A's 0x800F452C (112 B, maspsx=epilogue, second attempt) -- a sibling of its own
0x800F4098, matched with the mid-function-branch-target lever applied first try.
Worker A asked FIVE TIMES for the lui-page rule to be prominent, having made the mistake five
times. It is now cookbook 154 and the FIRST of the five things the workflow doc tells a new
worker, with all five instances tabulated. The fifth instance added the missing half: the
immediate is a SIGNED DISPLACEMENT, and the signature of the error is 'a residual of ONE
differing byte whose only difference is a lui or addiu immediate'.
Also confirmed from worker A's side: the shape A/B correction works -- 0x800F452C is a shape-B
row and closed with the single token, so the tail filter is obsolete and A has dropped it.
Worker A's 0x800F4098 and worker D's 0x800FB54C (104 B, first attempt, maspsx=epilogue).
151: the 2^k-1 add-back rule is CONFIRMED on two independent divisors -- worker C derived it
from 63 (0x800FEE3C) and worker D found it again on 127 (0x800FB54C, magic 0x81024409). Same
structure, two divisors, so finding 67's decision table is complete and not hypothesised.
152: FIVE finders each produced bodies over worker D's 20 -- redundancy rank 6, size rank 5,
adjacency 4, epilogue class 2, constant-division census 1, family 1. No single finder dominates.
This broadens finding 109: 'five different finders each produced bodies, and the price was set
by the LEVER, not the finder.' The tools cover different parts of the population, so keep every
finder running rather than consolidating onto the current best.
153: a saved register can force a local to be SMALLER than the data written through it, and
enlarging it to fix that breaks the frame.
Worker E's first four claims (0x80028CE0, 0x80028C34, 0x8009D798, 0x80091674).
147 IS A CORRECTION TO THE COORDINATOR'S OWN BROADCAST. I sent workers a 120-row epilogue list
selected by scanning the ORIGINAL's tail. Worker E showed that is not sufficient: 0x80091674 has
that exact tail shape but cc1 FILLS the slot itself, so the default toolchain matches (104 B) and
--fill-epilogue BREAKS it (100 B LENGTH-MISMATCH). The mode must be decided per row by compiling
and reading the CANDIDATE's tail. Diagnostic: 4 bytes SHORT with the mode on means the candidate
was already correct.
148: order and allocation are SEPARABLE -- the la position and the addu destination are two
independent facts, and the fix is to split the base into its own named local. Three spellings
with exact results.
149: a body 4-8 bytes SHORT with MISSING load-delay nops means a later independent block must
move earlier in the source -- the opposite tell from finding 100.
150: worker E hit a RACE DUPLICATE on 0x80100038 (free.sh read a registry 6 rows stale),
restored HEAD's file verbatim, re-verified it, dropped the row, and reported the incident.
Nothing corrupted because both spellings match -- but it proves the protocol's window is real.
Worker D's 0x800FAF84 (104 B), its first maspsx=epilogue match.
144: THE EPILOGUE CLASS NEEDS ONLY ONE TOKEN. Worker A asked for a second one; it does not need
it. The 120 rows split into two shapes -- A) lw $31 immediately before the release, which needs
the release moved AND a nop inserted after lw $31; B) other loads in between, where the release
moves and the trailing nop is DROPPED. My first implementation did A only and left every B row
4 bytes long. Verified on both: 0x800FFBEC (80/0/MATCH) and 0x800F44D0 (92/0/MATCH, a row worker
A had released as unfixable).
145: read the frame arithmetic and the saved-register offsets TOGETHER -- worker D's local had to
be 8 bytes not 12 because the saved s0 sits at sp+24 and the callee writes through sp+16. Third
instance of the size family, first where the constraint came from a saved register.
146: the SAME expression at two divisors produces two unrelated code shapes (/64 branchy bias vs
/63 add-back magic), which is why worker D's divisor sweep missed it.
Worker A released this row as 'the three-load-with-nop shape that the swap cannot fix' and
requested a --no-load-delay-nop token. It does not need one: the corrected transform already
handles it, by DROPPING the trailing nop for shape B rather than moving it. Verified: 92 bytes,
differing_bytes=0, MATCH, with maspsx=epilogue. It was 96 bytes before the fix.
So the epilogue class does NOT need a second token -- it needed the transform to distinguish the
two shapes, which worker A's report 24 is what revealed. The 120 rows should now be attemptable
with maspsx=epilogue alone.
Worker A's two epilogue-class rows (0x800F42AC 96 B, 0x80100038 104 B), both carrying the
maspsx=epilogue token -- the first rows closed through the new mode.
AND THE TRANSFORM IS NOW CORRECT FOR BOTH SHAPES, which worker A's report 24 showed was
necessary. The 120 rows split:
A) lw $31 IMMEDIATELY before the release -> the release moves into the slot AND a nop must be
inserted after lw $31, or j $31 lands in its load-delay slot. 0x800FFBEC.
B) other loads between lw $31 and the release -> the release moves into the slot and the
trailing nop is DROPPED; no load-delay nop is needed. Worker A's 0x800F44D0.
My first implementation did A only and left every B row 4 bytes long. Both are handled now, and
the discriminator is whether the jump's own register was loaded immediately before the release.
A BUG WORTH RECORDING: reading out[-1] to find that preceding instruction saw maspsx's own
'#nop # DEBUG: ...' comment instead of the lw, silently producing the shape-B answer for a
shape-A row and turning a MATCH back into a LENGTH-MISMATCH. The scan now skips comments.
Finding 122 declared GTE rows a blocked class needing a harness token. The gap was smaller than
it looked: the header already covered $0-$7 and $13-$31 but NOT $8-$12, which is where
IR1/IR2/IR3 ($9/$10/$11) live, and it had no macro for the GTE command instruction itself.
Added gte_ldIR1/2/3, gte_stIR1/2/3, gte_ldMAC1/2/3, gte_stMAC1/2/3, and a GENERIC gte_cmd(field)
so any command whose 25-bit field has been read off the original can be emitted without first
earning a semantic name.
VERIFIED, NOT ASSUMED: the field values are confirmed by counting them over the whole binary
(0x486012 x51, 0x49E012 x11, 0x41E012 x4) and a test compilation emits gte_cmd(0x486012) ->
cop2 0x486012 -> 0x4A486012, exactly the original's word.
Five rows were blocked on this across three workers, all with complete or partial derivations.
The disposition matters as much as the token: three workers independently reached 'no amount of
spelling will find this' and declared it BLOCKED rather than grinding -- which is what made the
gap visible as a TOOLING problem instead of three separate source puzzles.
141: worker D's own reads_unsaved_saved scanner flagged 0x80103434, D read the row, and the
verdict is NOT a fragment -- the detector treats op 0x12 (COP2) as a GPR use, which for
ctc2/mtc2/mfc2 is wrong. Every COP2 row in that flag list is suspect. That is the THIRD
independent instance of the same lesson: every shape heuristic in this project has needed a
worker to read a flagged row and say 'this is legal, and here is why'.
142: worker C's 2^k-1 add-back rule CLOSES worker D's open question on 0x82082083. D recorded it
as a bounded negative with a named direction; C recognised the answer from an unrelated row.
The third time a worker's finding closed another worker's negative -- the strongest argument for
recording a named open question rather than a vague failure.
Worker A's two: the addition operand-order row (a1[i]+a0[i] vs a0[i]+a1[i] -- same length,
16 bytes apart, because cc1 evaluates the right-hand operand first) and the unconditional
p[0]=0 that lands in a branch delay slot.
Worker D's two: 0x8009B56C closed on cookbook 43 trigger 1 after D had nearly written the row
off, and 0x80018458.
Finding 84 named the transform; it is now implemented and 0x800FFBEC matches (80 B, 0 differing)
where it was 6 differing bytes without it.
IT IS A SWAP, NOT A MOVE, and getting that wrong cost one implementation: the candidate is
lw $31,16(sp) / addiu sp,sp,24 / jr $31 / nop and the original is lw $31 / nop / jr $31 /
addiu sp,sp,24 -- SAME instruction count, two words swapped. My first version moved the release
after the jump and dropped the nop, producing 3 instructions instead of 4 and turning an 80-byte
row into a 76-byte LENGTH-MISMATCH. A 'small mechanical transform' still has to be checked
against the bytes.
SCALE: 120 unclaimed rows have the filled epilogue in the ORIGINAL (scanned every worklist row's
tail for jr $31 followed by a positive addiu sp,sp,N). They are mostly SMALL -- 76, 76, 80, 92,
92, 96, 104 B -- so this is a large class of cheap rows that were blocked on a HARNESS GAP rather
than on source shape. 770 other rows have the unfilled shape and need nothing.
The tracked patch is regenerated and verified to reproduce both modified maspsx files from the
pristine checkout.
Worker D's 0x80025A2C (176 B, 2 spellings) -- the EXACT MIRROR of its own 0x80025ADC.
131 is now THREE instances with an identical mechanism (88->96, 48->56, 56->64, always a clean
8-byte frame deficit with everything else identical, always fixed by an unreferenced array
local). It is a rule with a one-line fix, not a trick.
138: a MIRROR PAIR -- the two rows are inverse conversions, adjacent in the image, both callees
of 0x80031BBC. Finding one made the other a two-spelling row. sf3_family cannot see this relation
because inverses are not similar: their opcode histograms DIFFER (one stores where the other
loads). A pair of inverse functions has a STRUCTURAL signature, not a similarity one.
Worker A's 0x80036F70 (460 B, first attempt, family score 1.000 AND adjacent to its own
0x80036DA4). Its family run finished 7 for 7 with five first-spelling matches.
136: worker A CALIBRATED the family tool. It checked the two 0.97-scoring entries and NEITHER
shares its sibling's body at all -- one is a table-allocation routine, the other a summing
loop. '1.000 is the useful band; below ~0.99 the histogram is matching common idioms, not
bodies.' That is the same false-positive mode as the redundancy ranker (finding 110). The
default threshold is now 0.99.
137: a family's signature can be a CONSTANT TRIPLE -- worker A's 0x80036F70 differs from its
sibling only in six constants, whose signature is (A, A+12, A-58). Searchable in a way no
similarity metric can be, because the shapes are identical and only the immediates differ.
132: worker B found the SECOND independent instance of 'when the original's short path is the
fall-through, INVERT the condition' (30 bytes of layout on 0x800A8984; the same shape as its
own 0x800FCA90). That promotes it from a heuristic to a rule with a stated direction -- and it
is the opposite of the usual instinct to write the guard as an early-exit.
133: worker A's one-byte family residual was a DECLARATION -- 'int i' emits slt where 'unsigned
int i' emits the original's sltiu. The family transfers the SHAPE and the LEVERS; the
declarations must still be re-derived.
134: a family hit is also a hint about the CALLEE -- the relation crosses the call graph, and
sf3_family does not model it. Two of worker A's family rows call rows that are themselves
unclaimed with the same object layout.
135: worker A predicted a fourth family member by pattern; I scanned all 1046 unclaimed rows for
the predicted bases and ZERO reference them. The family has exactly three members. A predicted
member that does not exist is worth recording so nobody re-derives the search.
Worker A closed 0x800259A0, 0x80012918, 0x8006B2D4, 0x8003022C and 0x800506E4 -- every one a
sibling found by tools/sf3_family, which was built an hour ago from worker D's insight that
'the finder varies, the price does not' and therefore families should be SEARCHED for rather
than waited for.
That is the tool's first harvest and it is 5 bodies from one list. The family scores were
1.000/1.000/1.000/1.000/0.998 -- exact opcode-histogram and size matches against rows worker A
had already matched, so the levers transferred unchanged.
Records the cycle: five defects in coordinator-written rules (all found by workers following
them), the merge-flow defect the coordinator inflicted on itself, four new tools, the central
finding confirmed from both directions, the ranker's known blind spot with two failed proxies,
adjacency at 9-for-9, and the open items (GTE token, the post-pass, the solved 0x82082083).
Worker D's 0x80025ADC (136 B). Partitions re-interleaved 5 ways because workers B and C
are both at ~94% context and effectively exhausted, leaving 2 active workers against 45
remaining bodies. A fifth worker restores capacity.
130: worker D found the ranker's top is SIZE-BIASED -- redundancy and size are correlated because
a bigger body has more chances to repeat a 3-gram. Its top was 3288 B at 0.91 while the rows it
was matching sat at 0.64-0.74 in the small tail. The metric is not wrong about rows of equal
size; the raw score just cannot be compared across bands. Since the milestone counts BODIES, a
3288 B row and a 248 B row are worth one body each and the large one costs many times the
context. Worker D's effective filter was redundancy PER UNIT SIZE, now approximated by
--max-size, which is documented in the tool.
131: an unreferenced ARRAY local homes but an unreferenced SCALAR does not -- second independent
instance, identical mechanism, so it is a rule: when the frame is a clean multiple of 8 bytes
short and everything else is identical, add an unreferenced array local of that size.
Worker D's finding 125 said three members of one family were found by three different means
and 'the finder varies, the price does not', concluding that families should be SEARCHED FOR
explicitly rather than waited for. This implements that: for every unmatched worklist row,
find the already-matched row with the highest similarity, where similarity is an
opcode-histogram cosine (registers erased, nops dropped, per finding 120) multiplied by the
size ratio so a shared multiset at a different scale does not count.
A BUG WORTH RECORDING: the first version read config/match_worklist.tsv with the regions
column layout, so column 0 (the RANK) was read as the address. It returned ZERO candidates at
every threshold, which is what exposed it -- a silently wrong address yields no matches rather
than an error. Both layouts are now parsed by named functions with the offset documented.
Worker C's 0x8009F4B4 (248 B) and 0x80068874 (156 B), both first attempt.
127 CLOSES WORKER D'S OPEN QUESTION. D left 0x800FEE3C's magic 0x82082083 unexplained; worker C
solved it and the answer is a general rule: the divisor 63 is of the form 2^k-1, which is why
cc1 uses that magic with an ADD-BACK (mfhi; addu; sra 5) instead of a plain shift. An add-back
magic is the tell for a 2^k-1 divisor, NOT for a large one. Finding 67's decision procedure is
now complete: no mflo -> constant division D = 2^(32+s)/M; mfhi+addu+sra -> a 2^k-1 divisor;
mfhi AND mflo -> a genuine 64-bit multiply.
128: worker C classified a division-by-constant row on decode WITHOUT attempting it, because
'every division expression has several equally-plausible spellings, so it is idiom-redundant by
construction'. That characterises the ranker's false-positive class from the SOURCE side for the
first time -- exactly the class finding 110 showed cannot be separated by operand comparison.
129: adjacency is now 9-for-9 across three workers (A 3/3, C 5/5, D 1/1).
Worker A found that last-row-wins makes an unconditional 'released' from a worker who never
held the row WRONGLY FREE IT from whoever does. A found two addresses held as wip by worker C
and correctly re-appended C's wip row rather than its own released -- had it not, C's rows
would have been silently released and both workers could have started them.
New rule: only append 'released' for an address you yourself appended 'wip' for. Order is
check (range-aware) -> if free append wip -> work -> append your own released.
Second finding from the same episode: the roving list I gave worker A deliberately overlapped
the other partitions and its first two picks were both already held. A shared queue is only
worth it when the work is NOT already partitioned. Once the partitions were re-ranked on
global redundancy, the roving list was strictly worse than a worker's own slice, so it is
retired. That is the fifth defect found in a coordinator-written rule this phase.
Worker D's 0x8009F890 (248 B, first attempt) -- the THIRD member of a family it had already
matched twice, and the three were found by three different means: the size ranker, adjacency,
and the rebuilt global-redundancy rank. 'The finder varies, the price does not.' The
operational conclusion: families should be SEARCHED for explicitly rather than waited for.
126: worker D withdrew a named direction after checking its own notes and finding it had
already been tested and made the row worse. It declined a fifth spelling on a falsified lever
and corrected the row's record to the honest state. A direction that has been tested and
failed must be struck, or the next worker inherits a false lead.
123: a 'clamp to zero' written as a BITWISE MASK (x & (x >> 31)) compiles branchlessly while
the same thing as a TERNARY branches -- the opposite of the intuition. Worker B's MIN half now
matches exactly with the bitwise form. The MAX half (branchless slt/negu/and, i.e. x & -(x>0))
remains UNREACHED by any ternary or bitwise spelling, with a plausible SDK min/max macro as the
source. Recorded as a named open question.
124: worker B found that the partition design CONFOUNDED the size/redundancy measurement. The
partitions were rank-interleaved by the original worklist order (tier, size, address), not by
redundancy, so 'my <=200 B hit rate was high because my slice happened to be redundancy-rich,
not because small rows are inherently easy'. The within-worker evidence for 'cost is redundancy'
stands (A: 548 B first-attempt vs 176 B nine failures), but any CROSS-worker band comparison was
measuring the partition, not the rows. Fixed by re-partitioning on global redundancy. General
lesson: when work is divided among workers, any per-worker statistic is contaminated by the
division.
Worker A's 0x800689DC and worker C's 0x8009F4B4.
120: worker B's justification for why the ranker works -- 'the allocator makes copies
non-identical, so OPCODE repetition survives while WORD repetition does not'. That is exactly
why finding 110's full-word metric failed and why the opcode metric works. A repeated source
block produces the same opcodes with different registers; requiring operands to match destroys
the signal rather than sharpening it.
121: the filled-delay-slot class has TWO sub-cases with DIFFERENT fixes -- reorg fills the slot
(source-shape hunt) versus maspsx mode changing WHICH instruction lands in the slot (a harness
token choice). Same diagnostic, different remedy. Check whether toggling maspsx changes the
fill before hunting a source shape.
122: NEW BLOCKED CLASS -- a GTE coprocessor body needs a harness token, not more spellings.
Worker B's 0x8001FAFC reads mfc2 $12/$13/$14 and branches on t7/s6 which are NOT the o32
argument registers, so the inputs arrive through a non-standard convention. Team rule: if a
body contains mfc2/mtc2, do not spend spellings on it -- these are tooling-blocked rows to be
worked as a batch once a token exists.
Worker D's 0x80106AA8 (136 B, first attempt) -- found by the REDUNDANCY filter, not
adjacency, which is the first row where the ranker did the finding alone. Eight stores
through four global pointers, each re-materialised per store.
Cookbook 105's dial now has THREE measured settings: per statement (0x8006BC74 46x and
0x80106AA8 8x, both matched), once per block (matched), once per function (does not match).
So per-statement re-reads are the NORMAL shape, not an extreme.
119: worker D ran the fragment check, called 0x80058BA0 a confirmed fragment, then
SELF-CORRECTED -- it is legal, because in o32 a frameless leaf may both read and write the
caller's outgoing argument area (sp+0..sp+31). All three of D's suspects are legal. Worker B
found the read side, worker D the write side, and both had to read the row to do it: a
heuristic keyed on shape must state its exclusions, and only the worker reading the row can
find them.
116: worker A's hypothesis that the ranker cannot see LIVE RANGES (its conversion rate went to
zero over three consecutive rows, all failing on register allocation with the structure fully
confirmed). I tried to make it computable -- callee-saved registers saved and s-registers per
call -- and the proxy FAILED: matches span 0.00-2.00 s/call and failures span 0.17-1.33, with
a matched row having the HIGHEST s-register count and another matched row having 16 calls and
0 saved registers. The hypothesis is not refuted; the proxy is just not a good
operationalisation of 'live range'. Recorded as an open axis.
117: worker C's new signature -- correct control flow, +3 instructions, all three being j/jr
delay slots. DISTINCT from a nesting error (both give a small residual, but nesting moves
branch displacements while this moves instructions ACROSS a jump). Belongs with the post-pass
family.
118: call COUNT does not separate a real block from a call chain -- worker C's matched
0x80058CE8 is four loops that each call, and its false positive has zero calls. Fails in both
directions; recorded so nobody re-derives it.
Worker B's 0x80050CA8 (120 B, first attempt).
Lever: the status word is masked by TWO separate statements (&= -3; &= -5;), and the original
emits one load, two ands against two different constants, one store. Combining the masks
folds to a single and and LOSES an instruction -- the same principle as finding 81 (a slot
stored twice is two statements) applied to read-modify-write. Companion: the status load is
hoisted above nine halfword clears, so the clears' source order is only observable through
the store order.
The first version of the fragment check fired on any nonzero sp offset and flagged 2 of the
555 registered regions. Worker B read its own hit (0x800B704C) before committing and showed
it is a LEGAL FRAMELESS LEAF WITH EIGHT ARGUMENTS: in o32 the callee's sp is unchanged at
entry, so sp+16..sp+28 IS the caller's outgoing area -- arguments 4-7 -- and reading it
before any addiu sp,sp,-N is exactly what a frameless >4-argument leaf looks like. Its
evidence: exactly one jr ra, zero jal, zero addiu sp,sp,-N, zero sw ra/lw ra across all 324
bytes, and no callee-saved register touched.
Excluding the incoming argument area (sp+0..sp+31) and flagging only a negative offset or an
offset beyond the 8-argument area makes the check DISJOINT:
registered regions flagged: 2 of 555 -> 0 of 555
suspects across 4 partitions: 5 -> 1
The one remaining suspect is worker A's 0x800C3490.
This is the cleanest example in the phase of a worker ADJUDICATING a tool's output rather
than obeying it -- B was explicitly told 'advisory, do not skip', read the row anyway, and
its adjudication turned a noisy heuristic into a precise one.
Christopher Williams