|
|
|
|
@@ -0,0 +1,418 @@
|
|
|
|
|
# rx35 compiler correction and allocator evidence
|
|
|
|
|
|
|
|
|
|
Worktree sol-high, branch agent/w1005/rx35, HEAD and fetched origin/main
|
|
|
|
|
38ade13d6fb4320b5779d1e6df4b9250a018b7d8. Existing rx4 logs are untracked
|
|
|
|
|
and retained. This run owns configure.py and eZiText source only.
|
|
|
|
|
|
|
|
|
|
Read AGENTS.md, f-round.md, common.md, levers.md including 20-22, and the
|
|
|
|
|
mwdbg README before source work. Applied unslop and writing-for-agents.
|
|
|
|
|
The task authorizes the ZI8 compiler correction and Matching flips despite
|
|
|
|
|
the older common.md matching-only instruction.
|
|
|
|
|
|
|
|
|
|
Acceptance requires every baseline-exact eZiText function to remain exact,
|
|
|
|
|
unchanged or improved data, honest equivalent C, fresh object evidence,
|
|
|
|
|
symorder and doldiff for linking, full gate, and DOL SHA1
|
|
|
|
|
26116613f624061ba99c8d1a299aaa6efa85670d. A compiler correction and its source
|
|
|
|
|
retuning must be committed together only after that gate passes.
|
|
|
|
|
|
|
|
|
|
The installed mwdbg pins GC/3.0a5.2 and rejects GC/3.0a5 before execution.
|
|
|
|
|
Investigating a local debugger mapping under build/rx35; shared tools remain
|
|
|
|
|
unchanged. Captures will need independent whole-object identity verification.
|
|
|
|
|
The ccsweep2 copy changes only its output directory to keep trial objects here.
|
|
|
|
|
|
|
|
|
|
Prior rx4 evidence was read for workflow and data-section constraints. Its
|
|
|
|
|
GC/3.0a5.2 allocator conclusions are not assumed to hold for GC/3.0a5.
|
|
|
|
|
|
|
|
|
|
Fresh compiler sweep: EZTX is 7/7 instruction-exact under both versions, so
|
|
|
|
|
EZTX stays on GC/3.0a5.2. ZI8 GC/3.0a5 immediately matches both zi81key
|
|
|
|
|
spelling helpers, two zi8cgetc helpers, both zi8getc2 helpers, Zi8GetPyFinal,
|
|
|
|
|
ZiDAWGgetCHARattribute, and three zkokeyp helpers. The fresh baseline already
|
|
|
|
|
has ZiDAWGGetGraphInfo exact. All eZiText string pools are identical.
|
|
|
|
|
|
|
|
|
|
Regression target maps: Zi8MatchROMdata1 wants elements/count/output/capacity/
|
|
|
|
|
mode/acceptPrefix in r20/r21/r22/r23/r24/r25 and the table address in r26.
|
|
|
|
|
GC/3.0a5 spills the current unused integer tableAddress to the stack and shifts
|
|
|
|
|
those saved registers. Both exclusion functions want ending r27, beginning
|
|
|
|
|
r28; current code reverses those. Zi8GetBpmfPhonetic wants count r26, value r27;
|
|
|
|
|
current code reverses those. Zi8_81480224 wants candidate r26, position r27,
|
|
|
|
|
word-priority evaluated first, and index-first address additions.
|
|
|
|
|
|
|
|
|
|
The local debugger port maps 24 hooks by unique relocation-masked instruction
|
|
|
|
|
sequences, globals by corresponding instruction operands, and all 496 opcode
|
|
|
|
|
names by exact table comparison. The after-scheduling hook has the same call
|
|
|
|
|
sequence with one x86 local register changed. Runtime capture and whole-object
|
|
|
|
|
validation remain required before treating the port as evidence.
|
|
|
|
|
|
|
|
|
|
Manual source trials restored three exact regressions. Splitting tableAddress
|
|
|
|
|
assignment from the groupPtr store restores Zi8MatchROMdata1 to 128/128,
|
|
|
|
|
100%, data 120/120; nesting a pointer-typed assignment did not. Using a
|
|
|
|
|
ziWChar pointer for beginning and removing the void-pointer indexing cast
|
|
|
|
|
restores both exclusions to 51/51 and 100%. Reordering declarations or
|
|
|
|
|
splitting the postfix length decrement did not.
|
|
|
|
|
|
|
|
|
|
Zi8GetBpmfPhonetic restores 309/309 and 100% by combining each count decrement
|
|
|
|
|
with its following zero test. These are separate full expressions in the
|
|
|
|
|
same sequence; the tested value remains the decremented byte. Per-site
|
|
|
|
|
expanded subtraction, declaration order, count-copy and full-width local
|
|
|
|
|
controls did not match. No caller or public type changes are retained.
|
|
|
|
|
|
|
|
|
|
With the split DAWG assignment retained, a fresh link already restores the
|
|
|
|
|
required DOL SHA1. doldiff reports zero differing bytes. The other regression
|
|
|
|
|
fixes and new Matching flips still need their final gate.
|
|
|
|
|
|
|
|
|
|
New exact functions under GC/3.0a5: Zi8_8148302C uses a halfword character
|
|
|
|
|
value declared after its table count, 59/59 and 100%; changing the type alone
|
|
|
|
|
did not fix r27/r28. Zi8getKeyLayout chains keyIndex = tableCount = 0, 182/182
|
|
|
|
|
and 100%; this moves tableCount from r27 to target r26 and language from r26
|
|
|
|
|
to target r27. Splitting call assignments and declaration-order controls did
|
|
|
|
|
not move that pair. The named initialization is semantically unchanged.
|
|
|
|
|
|
|
|
|
|
The user-word helper improves from the old main 99.52586 to 99.82758 with
|
|
|
|
|
position declared before candidate and the word priority tested first.
|
|
|
|
|
Its remaining two differences commute the operands of address additions.
|
|
|
|
|
Byte-offset and flexible-array controls did not match; none is retained.
|
|
|
|
|
|
|
|
|
|
All three initial Matching flips pass the linked DOL hash and doldiff with
|
|
|
|
|
zero bytes different. symorder's only discrepancies were anonymous table and
|
|
|
|
|
exception labels. A second comparison checked every offset, size and named
|
|
|
|
|
symbol after normalizing only those anonymous names; all three passed.
|
|
|
|
|
|
|
|
|
|
Fresh comparison proves the large zi81key functions, Zi8AlphaGetCandidates,
|
|
|
|
|
zi8InternalGetZH and Zi8PrepareMatch have identical instruction streams under
|
|
|
|
|
GC/3.0a5 and GC/3.0a5.2. Their earlier allocator hypotheses are therefore
|
|
|
|
|
being rechecked against new captures, not assumed fixed by the version change.
|
|
|
|
|
|
|
|
|
|
Zi8MatchOEMdata is now 192/192 and 100%. Combining all three index updates
|
|
|
|
|
with their consumers fixes the saved-register cycle: initialize index in its
|
|
|
|
|
limit check, test ++index directly, and store ++index into oemIdx. One or two
|
|
|
|
|
combinations reach 99.40104 with only index/length swapped; all three put
|
|
|
|
|
index r26, length r27, pattern r28 and work r29 as retail requires. Splitting
|
|
|
|
|
the capacity decrement is a rejected control. Every index update and read
|
|
|
|
|
remains in its original execution order.
|
|
|
|
|
|
|
|
|
|
Data-preserving source controls were run on the larger functions. The three
|
|
|
|
|
initial spelling-control tags accidentally used absent source spellings and
|
|
|
|
|
are no-ops, excluded from attempt counts. Corrected postfix, compound and
|
|
|
|
|
expanded-index forms were then compiled. New a5 measurements reproduce the
|
|
|
|
|
old 1697/1701 instruction behavior. ZH const-input changes add nine
|
|
|
|
|
instructions and lose data; native type and compound-mask controls do not
|
|
|
|
|
move output. Alpha initialization gains a fuzzy fraction but adds two
|
|
|
|
|
instructions and loses data. Prepare's chained first-element assignment,
|
|
|
|
|
explicit component-byte cast and split phonetic zero stores were tested
|
|
|
|
|
alone and in all seven combinations; none preserves the full data credit.
|
|
|
|
|
|
|
|
|
|
The shared mwdbg requests stayed queued for over 20 minutes. They were
|
|
|
|
|
canceled before capture. A private copy of retrowin32 now listens on localhost
|
|
|
|
|
19035, with its own worktree lock. Only the two-byte TCP port immediate changed
|
|
|
|
|
in that emulator copy; compiler executables and compiler flags are unchanged.
|
|
|
|
|
The shared source-search CPU limiter is unchanged. Port patch hashes and the
|
|
|
|
|
verified x86 instruction are saved in build/rx35/emulator-port.json.
|
|
|
|
|
|
|
|
|
|
First a5 debugger run exposed a real layout difference: the function link-name
|
|
|
|
|
field is +0x40, not a5.2's +0x44. The a5 LinkName routine at 0x556400 and the
|
|
|
|
|
runtime object both prove +0x40. The mapping was corrected before using any
|
|
|
|
|
allocator evidence. Failed diagnostics are retained under build/rx35/private-*.
|
|
|
|
|
|
|
|
|
|
The corrected a5 debugger capture completed. Its alpha object is byte-identical
|
|
|
|
|
to an independent wibo compile. The validator checked reversed simplify order,
|
|
|
|
|
assignment events, representative maps and every PCode register rewrite with
|
|
|
|
|
zero mismatches. This validates the local port for this capture, not every
|
|
|
|
|
unsupported debugger feature.
|
|
|
|
|
|
|
|
|
|
The exclusion dump exposes an important distinction: ending r28 and beginning
|
|
|
|
|
r27 are already physical in backend-00 before graph coloring. The graph pass
|
|
|
|
|
only assigns temporaries r32-r41; r39 merges into call-result r3, and the
|
|
|
|
|
parameter moves are rejected for interference. Therefore the surviving
|
|
|
|
|
callee-saved swap originates in the earlier no-opt variable assignment, not
|
|
|
|
|
in simplify degree or graph coalescing. A typed-beginning capture will confirm
|
|
|
|
|
the changed preassignment. The earlier heuristic about graph priority alone
|
|
|
|
|
does not explain these named-register swaps.
|
|
|
|
|
|
|
|
|
|
The a5 no-opt selector was traced at 0x46ad60. It scans parameter and local
|
|
|
|
|
objects, chooses the greatest reference weight, and replaces a choice on an
|
|
|
|
|
equal weight. Named locals are preassigned before the graph pass begins.
|
|
|
|
|
Zi8ITspecialExclusion: void beginning has weight 2, below ending 3, so ending
|
|
|
|
|
gets r28 and beginning r27. Making beginning ziWChar* raises it to weight 3;
|
|
|
|
|
the local wins the parameter tie and gets target r28, leaving ending r27.
|
|
|
|
|
The graph priority and call-result coalescing stay unchanged.
|
|
|
|
|
Zi8getKeyLayout: separate tableCount=0; keyIndex=tableCount makes tableCount
|
|
|
|
|
weight 7, tied with language, and the local wins r27. Chaining their zero
|
|
|
|
|
initialization removes the extra tableCount read: weight 6, r26. Language 7
|
|
|
|
|
now takes target r27. Both mwdbg objects equal independent wibo objects.
|
|
|
|
|
The remaining UWD and Korean-table address additions retain opposite commuted
|
|
|
|
|
operands after typed-member, array, integer-address and const controls. No
|
|
|
|
|
integer-address or byte-offset experiment is retained. Three 1200-second
|
|
|
|
|
search seeds (49,149,249) now also cover these two functions, using the shared
|
|
|
|
|
24-slot limiter and immutable source snapshots inside this worktree.
|
|
|
|
|
|
|
|
|
|
Fresh no-opt trace confirms the DAWG repair: nested tableAddress assignment
|
|
|
|
|
has weight 1 and remains on the stack. Splitting the groupPtr assignment
|
|
|
|
|
raises tableAddress to weight 2, takes target r26, and shifts six weight-2
|
|
|
|
|
parameters down one register. The object is identical to an independent build.
|
|
|
|
|
OEM index weight falls 12 -> 9 when its three updates feed their consumers;
|
|
|
|
|
work/pattern stay weight 12 and length stays 10. The resulting assignments
|
|
|
|
|
are work r29, pattern r28, length r27, index r26, exactly matching retail.
|
|
|
|
|
|
|
|
|
|
Zi8_814834AC target maps i=r28, tableA load=r4, scaled byte index=r3.
|
|
|
|
|
The source pre-color ADD is v115=v112+v114: tableA v112 (degree 11,
|
|
|
|
|
priority index 80) resolves r4; offset v114 (degree 9, priority 78) resolves
|
|
|
|
|
r3; sum v115 (degree 10, priority 77) resolves r3. The second occurrence is
|
|
|
|
|
v122=v119+v121 with the same assignments. These colors already match retail;
|
|
|
|
|
the ADD operand order differs before coloring. This is an instruction-
|
|
|
|
|
selection difference, not a saved-register swap or coalescing failure.
|
|
|
|
|
Twenty further signed/unsigned address-expression controls did not fix it.
|
|
|
|
|
|
|
|
|
|
Prepare's combined element assignment, explicit halfword component-byte cast,
|
|
|
|
|
and separate phonetic zero stores remove all non-register structural changes
|
|
|
|
|
except three extra call-result copies (943 vs 940). This candidate loses data
|
|
|
|
|
and is not retained. Its mwdbg capture and three additional 1200-second seeds
|
|
|
|
|
349/449/549 are queued to investigate the actual coalescing differences.
|
|
|
|
|
A 64-shape spelling output-loop experiment did not reach exactness; the best
|
|
|
|
|
normal source shape is 1701 instructions and remains an unretained trial.
|
|
|
|
|
|
|
|
|
|
Debugger transport optimization: stop-local page reads initially missed a
|
|
|
|
|
cache invalidation after the debugger called LinkName, so those captures
|
|
|
|
|
reported empty names and failed before recording any result. The cache now
|
|
|
|
|
clears explicitly after every inferior resume and LinkName call. A repeated
|
|
|
|
|
51-instruction exclusion capture produced byte-identical regalloc JSON,
|
|
|
|
|
no-opt events, initial PCode and final PCode versus the uncached capture;
|
|
|
|
|
the independent object/PCode validator also passes. No failed capture is
|
|
|
|
|
included as measurement evidence. For large functions only the selected and
|
|
|
|
|
assigned no-opt hooks are enabled; graph priority/degree/coalescing hooks
|
|
|
|
|
remain complete. The redundant full candidate-scan ZH capture was interrupted
|
|
|
|
|
and requeued with that smaller no-opt hook set.
|
|
|
|
|
|
|
|
|
|
origin/main advanced externally to 8899a830 during this run. None of its
|
|
|
|
|
source edits touch eZiText. The rx35 branch remains on its assigned 38ade13d
|
|
|
|
|
baseline; no fetch, rebase, merge or edit outside this worktree was performed
|
|
|
|
|
while examining the new ref. Final regression comparisons use the recorded
|
|
|
|
|
38ade13d baseline and independent baseline-report.json snapshot.
|
|
|
|
|
|
|
|
|
|
Transport experiments enabled TCP_NODELAY on only the private emulator/GDB
|
|
|
|
|
connection. A temporary native shim changes socket options, never compiler
|
|
|
|
|
memory or flags; both the object and allocator output remain unchanged. The
|
|
|
|
|
full event trace is still too slow on the 10676-instruction ZH function.
|
|
|
|
|
A snapshot mode was therefore validated against the full exclusion capture:
|
|
|
|
|
actual compiler priority list, before-simplify graph, final graph, every
|
|
|
|
|
register rewrite and final PCode are identical. Large captures retain all
|
|
|
|
|
coalescing events and named-register assignments, but skip per-node simplify
|
|
|
|
|
and color-event breakpoints. Their validator explicitly marks those skipped
|
|
|
|
|
event checks null; graph/PCode and independent-object checks remain required.
|
|
|
|
|
The interrupted ZH event trace is partial evidence only, not a validated
|
|
|
|
|
capture. Its complete snapshot capture is queued on a separate private port.
|
|
|
|
|
|
|
|
|
|
All original source-search seeds finished. No candidate from seeds
|
|
|
|
|
49/149/249 produced another exact function. Korean state flattening into
|
|
|
|
|
normal scalar locals and a real key array was also tested: 682-684 versus
|
|
|
|
|
669 target instructions, 93.55-94.49%, data 40/180. None is retained.
|
|
|
|
|
|
|
|
|
|
Reproduction artifacts: rx35.mwdbg-a5.patch records the private a5 driver,
|
|
|
|
|
launcher and optional TCP shim relative to the shared tool versions whose
|
|
|
|
|
SHA256 values are in rx35.mwdbg-port.json. Apply it only to copies. The launcher
|
|
|
|
|
expects its private files under the current worktree's build/rx35 directory;
|
|
|
|
|
its default project is now the current directory. Shared debugger files,
|
|
|
|
|
compiler executables, and other worktrees were not edited.
|
|
|
|
|
|
|
|
|
|
Confirmed Bpmf no-opt weights: count 22 -> 15, value remains 21; value moves
|
|
|
|
|
r26 -> target r27 and count r27 -> target r26. Graph priority is descending
|
|
|
|
|
v147..v32, max simplify degree 16, with no accepted coalesces in either shape.
|
|
|
|
|
The count/value change is therefore entirely in named-register preassignment.
|
|
|
|
|
Korean key helper: count/value both weight 2. Their declaration/type order
|
|
|
|
|
changes the local tie order from value r28/count r27 to target count r28/value
|
|
|
|
|
r27; value's explicit halfword calculation is unchanged.
|
|
|
|
|
UWD: position/candidate both weight 7. Declaring position first changes their
|
|
|
|
|
local tie order from candidate r27/position r26 to target position r27/
|
|
|
|
|
candidate r26. Word is r28, work r31, current r30 and added r29 in both.
|
|
|
|
|
The retained priority comparison orders the reads as in retail. Two ADD
|
|
|
|
|
operand orders remain; no forced-register source was introduced.
|
|
|
|
|
|
|
|
|
|
Prepare structural candidate, measured graph evidence: extra copies at source
|
|
|
|
|
instruction 87, 646 and 654 are v62->r4, v103->r0 and v104->r4, while retail
|
|
|
|
|
stores r3 directly. Before simplification their degrees are 7, 6 and 7;
|
|
|
|
|
priority positions are 189, 150 and 149 (zero-based). All have r3 as an
|
|
|
|
|
interference neighbor. v62 is also the strokePrefix address in block 117,
|
|
|
|
|
live beside an index already coalesced to r3; v103/v104 are also the index
|
|
|
|
|
and recordMasks address in block 28, live beside the 0xff value coalesced to
|
|
|
|
|
r3. Thus unrelated uses of the same virtual temporary slots block merging
|
|
|
|
|
those call-result copies. Sixty-three unsigned-index, typed-constant and
|
|
|
|
|
commuted-OR source variants did not remove all three copies. Three further
|
|
|
|
|
annealing seeds 349/449/549 completed; only a tiny fuzzy-only candidate was
|
|
|
|
|
saved, still 943/940 with data loss. It is not retained.
|
|
|
|
|
|
|
|
|
|
Korean candidates: retail requires keyTable r28/work r27. Current no-opt
|
|
|
|
|
weights are keyTable 13/work 14, so the compiler assigns them r27/r28. The
|
|
|
|
|
other saved values match: param r31 (110), options r30 (40), wordNode r29
|
|
|
|
|
(27). Removing an identity pointer cast or changing its void view leaves
|
|
|
|
|
code unchanged. Flattening the pre-existing state makes keyTable miss the
|
|
|
|
|
no-opt eligibility threshold, adds stack loads, and regresses the function.
|
|
|
|
|
Reversing real local declarations repairs much of its stack layout, but
|
|
|
|
|
produces 683/669. Using a typed table gives 661/669 and loses data. None is
|
|
|
|
|
retained; no new carrier structure is added.
|
|
|
|
|
|
|
|
|
|
The larger functions retain their target saved-register roles. Spelling uses
|
|
|
|
|
tableIndex r31, params r30, character r29, work r28, options r27. Its 207-node
|
|
|
|
|
priority starts v259,v258,v257; 21 coalesces and 212 interference rejections
|
|
|
|
|
remain after the postfix/index trial. The trial changes degrees but keeps
|
|
|
|
|
1701/1700 instructions. Retail needs one fewer byte narrowing and a different
|
|
|
|
|
pair of support-call copies. Candidates uses params r31, index r30, work r29,
|
|
|
|
|
options r28; its 188-node priority starts v258,v257,v256, with 39 coalesces and
|
|
|
|
|
182 interference rejections. It remains 2405/2397 with three extra capacity
|
|
|
|
|
reloads and a net balance of call-result/index copies.
|
|
|
|
|
Alpha uses parameters r31, workData r30, index r29, wordCursor r28,
|
|
|
|
|
optionData r27. Chained dictionary initialization changes 171 -> 172 colored
|
|
|
|
|
temporaries and 55 -> 56 coalesces; degree graphs and code/data regress, so
|
|
|
|
|
it is rejected. ZH uses request r31, work r30 and options r29. Making its
|
|
|
|
|
input read-only leaves those named weights 352/277/99 unchanged but changes
|
|
|
|
|
206 -> 210 colored temporaries and 31 -> 27 coalesces, adding nine instructions.
|
|
|
|
|
The live ZH and Alpha differences include expression/store order as well as
|
|
|
|
|
register choices; neither is classified as a proven pure allocator tie.
|
|
|
|
|
All degree summaries for snapshot captures refer to the before-simplify graph.
|
|
|
|
|
Every PCode register rewrite checked in these captures agrees with the final
|
|
|
|
|
graph: Spelling 2979, Candidates 4499, Alpha 6545 and ZH 18821 operands.
|
|
|
|
|
|
|
|
|
|
Final clean gate: PASS, all 31 eZiText units included; zero baseline-exact
|
|
|
|
|
function drops, zero code/data/link regressions, zero forbidden additions,
|
|
|
|
|
zero readability warnings. The build was cleaned by gate.py before rebuilding.
|
|
|
|
|
The independent stored 38ade13d report also shows zero exact-function drops.
|
|
|
|
|
DOL SHA1: 26116613f624061ba99c8d1a299aaa6efa85670d.
|
|
|
|
|
doldiff: 0 differing bytes; both files 3867904 bytes.
|
|
|
|
|
New Matching units: zi8getc2, zidawg1, zi8match, zmtkey, zoemdata.
|
|
|
|
|
Their symbol order, offsets, sizes and named symbols agree after normalizing
|
|
|
|
|
only anonymous @ and jump-table names. Raw symorder logs are in build/rx35.
|
|
|
|
|
The actual linked DOL independently validates their placement.
|
|
|
|
|
|
|
|
|
|
EziText exact functions: 127 -> 141 / 149. Eight functions remain nonexact;
|
|
|
|
|
zi81key still accounts for the final 108 unmatched data bytes. The complete-
|
|
|
|
|
project checker reports DECOMPLETE_FAIL because code/data/link totals remain
|
|
|
|
|
incomplete. No project-completion claim is made. Global worker-baseline code:
|
|
|
|
|
94.00436% exact, 99.81261% fuzzy, 79.93427% linked; data 99.99410% matched.
|
|
|
|
|
Zi8GetKOcandidates has a disclosed fuzzy-only compiler-version drop,
|
|
|
|
|
98.87145 -> 98.72197; it was already nonexact and the required exact-name
|
|
|
|
|
regression gate passes. Full per-function before/after values and measures
|
|
|
|
|
are in rx35.results.json.
|
|
|
|
|
|
|
|
|
|
Completed searches:
|
|
|
|
|
Zi8MatchOEMdata-249.log: done 827 trials; best-nodrop energy -98.8802
|
|
|
|
|
Zi8MatchOEMdata-49.log: done 1382 trials; best-nodrop energy -98.8802
|
|
|
|
|
Zi8_81480224-149.log: done 1092 trials; best-nodrop energy -99.8276
|
|
|
|
|
zi8InternalGetZH-249.log: done 457 trials; best-nodrop energy -98.3363
|
|
|
|
|
Zi8Get1KeyPressCandidates-49.log: done 1139 trials; best-nodrop energy -80.4193
|
|
|
|
|
Zi8_81480224-49.log: done 1075 trials; best-nodrop energy -99.8276
|
|
|
|
|
Zi8PrepareMatch-249.log: done 792 trials; best-nodrop energy -96.9947
|
|
|
|
|
Zi8Get1KeyPressSpelling-249.log: done 713 trials; best-nodrop energy -97.5559
|
|
|
|
|
Zi8_814834AC-149.log: done 1066 trials; best-nodrop energy -99.9417
|
|
|
|
|
Zi8Get1KeyPressCandidates-149.log: done 1131 trials; best-nodrop energy -80.4193
|
|
|
|
|
Zi8_81480224-249.log: done 1072 trials; best-nodrop energy -99.8276
|
|
|
|
|
Zi8AlphaGetCandidates-49.log: done 1020 trials; best-nodrop energy -95.6485
|
|
|
|
|
zi8InternalGetZH-149.log: done 597 trials; best-nodrop energy -98.3363
|
|
|
|
|
Zi8Get1KeyPressCandidates-249.log: done 678 trials; best-nodrop energy -80.4193
|
|
|
|
|
Zi8_814834AC-49.log: done 1061 trials; best-nodrop energy -99.9417
|
|
|
|
|
Zi8MatchOEMdata-149.log: done 837 trials; best-nodrop energy -98.8802
|
|
|
|
|
Zi8PrepareMatch-49.log: done 1312 trials; best-nodrop energy -96.9947
|
|
|
|
|
zi8InternalGetZH-49.log: done 746 trials; best-nodrop energy -98.3363
|
|
|
|
|
Zi8GetKOcandidates-49.log: done 1284 trials; best-nodrop energy -98.722
|
|
|
|
|
Zi8getKeyLayout-49.log: done 1385 trials; best-nodrop energy -99.533
|
|
|
|
|
Zi8getKeyLayout-249.log: done 818 trials; best-nodrop energy -99.533
|
|
|
|
|
Zi8PrepareMatch-149.log: done 797 trials; best-nodrop energy -96.9947
|
|
|
|
|
Zi8GetKOcandidates-149.log: done 1300 trials; best-nodrop energy -98.722
|
|
|
|
|
Zi8Get1KeyPressSpelling-49.log: done 1164 trials; best-nodrop energy -97.5559
|
|
|
|
|
Zi8_814834AC-249.log: done 1051 trials; best-nodrop energy -99.9417
|
|
|
|
|
Zi8Get1KeyPressSpelling-149.log: done 1179 trials; best-nodrop energy -97.5559
|
|
|
|
|
Zi8GetKOcandidates-249.log: done 790 trials; best-nodrop energy -98.722
|
|
|
|
|
Zi8AlphaGetCandidates-249.log: done 850 trials; best-nodrop energy -95.6485
|
|
|
|
|
Zi8getKeyLayout-149.log: done 869 trials; best-nodrop energy -99.533
|
|
|
|
|
Zi8AlphaGetCandidates-149.log: done 654 trials; best-nodrop energy -95.6485
|
|
|
|
|
prepare-structural-search-349.log: done 1169 trials; best-nodrop energy -91.367
|
|
|
|
|
prepare-structural-search-449.log: done 1183 trials; best-nodrop energy -91.367
|
|
|
|
|
prepare-structural-search-549.log: done 1173 trials; best-nodrop energy -91.3681
|
|
|
|
|
|
|
|
|
|
Unit exact-function counts, recorded from the final fresh report:
|
|
|
|
|
zi81key: 5 -> 7 / 9
|
|
|
|
|
zi8alpha: 11 -> 11 / 12
|
|
|
|
|
zi8cgetc: 5 -> 7 / 8
|
|
|
|
|
zi8getc2: 15 -> 17 / 17
|
|
|
|
|
zi8match: 9 -> 10 / 10
|
|
|
|
|
zi8uwd: 3 -> 3 / 4
|
|
|
|
|
zidawg1: 5 -> 6 / 6
|
|
|
|
|
zkokeyp: 1 -> 5 / 7
|
|
|
|
|
zmtkey: 3 -> 4 / 4
|
|
|
|
|
zoemdata: 2 -> 3 / 3
|
|
|
|
|
zprepare: 0 -> 0 / 1
|
|
|
|
|
|
|
|
|
|
Final debugger validation:
|
|
|
|
|
All 40 completed captures match their independent wibo objects. The two large
|
|
|
|
|
zi8InternalGetZH captures contain three conditional branches whose targets
|
|
|
|
|
exceed the signed 16-bit branch range. Each becomes an inverted conditional
|
|
|
|
|
skip plus an unconditional branch; the apparent three-instruction PCode
|
|
|
|
|
alignment failure is fully explained. The supplemental validator checks the
|
|
|
|
|
inverted predicate (excluding its prediction hint), CR bit, eight-byte skip,
|
|
|
|
|
original out-of-range displacement and exact destination block. It then
|
|
|
|
|
checks all ordinary machine-register operands with those three expansions
|
|
|
|
|
accounted for: 8604 instructions / 16379 operands for the baseline and 8615 /
|
|
|
|
|
16401 for the rejected const-input trial. The original naive alignment result
|
|
|
|
|
is retained beside this proof in rx35.allocator.json.
|
|
|
|
|
|
|
|
|
|
The reproduction patch now includes the snapshot-aware validator and both
|
|
|
|
|
far-branch validation scripts. Run validate.py with a local build/tools/wibo,
|
|
|
|
|
then check-long-branches.py and validate-expanded-branches.py for the named
|
|
|
|
|
ZH captures. No shared debugger or other worktree was modified. All 141 exact
|
|
|
|
|
eZiText functions also passed fresh ctxdiff with identical instruction counts
|
|
|
|
|
and zero differences (build/rx35/final-ctxdiff.json). No source or configure
|
|
|
|
|
change was made after the clean all-unit gate.
|
|
|
|
|
|
|
|
|
|
## Landing repair: Zi8GetKOcandidates only
|
|
|
|
|
|
|
|
|
|
Resumed at fed3bdbb after the orchestrator rebased and removed bulky artifacts.
|
|
|
|
|
The new acceptance rule compares every function score, including fuzzy ones.
|
|
|
|
|
Required minimum: the old-compiler score 98.87145, with every other eZiText
|
|
|
|
|
function unchanged from this resumed branch. Only zkokeyp.c and this log are
|
|
|
|
|
in scope. All new captures, measurements and rejected trials stay in build/rx35r1.
|
|
|
|
|
|
|
|
|
|
Fresh source baseline: Zi8GetKOcandidates 98.72197, 669/669 instructions,
|
|
|
|
|
zkokeyp data 180/180. Pool identical. The other six function scores match the
|
|
|
|
|
accepted branch, including Zi8_814834AC at 99.94169.
|
|
|
|
|
|
|
|
|
|
Rejected controls: pointer/index association and signed byte views leave the
|
|
|
|
|
same output; arithmetic and typed-record subscripts usually change instruction
|
|
|
|
|
counts and lose data. Splitting output increments, changing condition forms,
|
|
|
|
|
removing redundant casts, and consolidating null-error handling did not meet
|
|
|
|
|
the score floor. A real five-byte array initializer reached 98.7728 but moved
|
|
|
|
|
stack objects; it is not retained. No carrier, forced register, redundant
|
|
|
|
|
operation or compiler change is retained.
|
|
|
|
|
|
|
|
|
|
Accepted source: read the first packed key flag through the byte table using
|
|
|
|
|
offsetof(ZiKoreanKeyEntry, keyBytes[4]). This names the same existing field
|
|
|
|
|
instead of casting the computed address to a record and then selecting it.
|
|
|
|
|
The row offset is still keyIndex*9; keyIndex is unsigned 16-bit and the 0xffff
|
|
|
|
|
sentinel still short-circuits before the read. The additional field offset is
|
|
|
|
|
four, so all arithmetic is within signed 32-bit range. The same unsigned byte
|
|
|
|
|
and bit 8 are tested, with unchanged calls, side effects and branch condition.
|
|
|
|
|
stddef.h supplies offsetof; no layout, symbol, flag or public type changed.
|
|
|
|
|
|
|
|
|
|
mwdbg confirms keyTable's reference weight rises 13 -> 14. It ties the work
|
|
|
|
|
parameter at 14 and wins the local/parameter tie: keyTable r28, work r27, as
|
|
|
|
|
retail requires. Param/options/wordNode remain r31/r30/r29 at weights
|
|
|
|
|
110/40/27. The 213-node GPR priority order and 14 successful coalesces remain;
|
|
|
|
|
52 candidate coalesces are rejected for interference. This is a named-register
|
|
|
|
|
preassignment repair, not a different graph coloring priority. The capture
|
|
|
|
|
matches an independent wibo object byte-for-byte. All 1224 allocation operand
|
|
|
|
|
rewrites and 1027 machine-register operands across 537 instructions pass.
|
|
|
|
|
|
|
|
|
|
Focused result: Zi8GetKOcandidates 98.72197 -> 98.998505, still 669/669.
|
|
|
|
|
All six other function scores are identical and data stays 180/180.
|
|
|
|
|
The clean eZiText gate and explicit per-function fuzzy comparisons follow.
|
|
|
|
|
|
|
|
|
|
The graph priority and coalesced representative pairs compare identical.
|
|
|
|
|
Before-simplify degree changes: v59 6->5, v58 8->7.
|
|
|
|
|
Focused ctxdiff improves from 103 to 71 differences, both 669 instructions.
|
|
|
|
|
There are 214 private compile records, including controls and baseline refreshes.
|
|
|
|
|
|
|
|
|
|
Final clean gate: GATE PASS across all 31 eZiText units, base 044ab571.
|
|
|
|
|
Baseline regressions, forbidden additions and readability warnings are all zero.
|
|
|
|
|
Fresh reports compare all 149 eZiText functions: only Zi8GetKOcandidates
|
|
|
|
|
changes, 98.72197 -> 98.998505. Every other function score is identical.
|
|
|
|
|
All eZiText scores meet or exceed the pre-switch 044ab571 baseline; the
|
|
|
|
|
old score for this function was 98.87145. No common function in the full
|
|
|
|
|
project report has a fuzzy drop against that baseline either.
|
|
|
|
|
DOL SHA1: 26116613f624061ba99c8d1a299aaa6efa85670d.
|
|
|
|
|
Only zkokeyp.c and this .md log are committed; artifacts remain under build/rx35r1.
|