Skip to content

CARVE ANSWER - N5SWAP fn 5 (MSWIN, "swap-in") driven into the real swapper (2026-08-17)

Milestone 14 of the CpuND500UC microword engine: fn 5 - the OTHER gate-bypass arm of the swapper's 29-way dispatch - issued as a real work order and executed DEEP into its worker on the real CONT-STORE-10611 microcode, plus the instruction batch this took (ADD3/SUB3/MUL3, MULAD, F/D DIV3, LOOPI, SHL, W GETBI, SFILL, BMOVE, l=:, ced=:, tutti) and the resolution of the CPGU/CWIP exit anomaly (TICK-MODEL item 78 -> item 83).

Evidence tags: PROVEN = read from the cited bytes/listing AND (where stated) executed on the microword engine. INFERRED = deduced, not fact. OPEN = unresolved.

Sources: - Microcode listing: E:\Dev\Repos\Ronny\ND110Compile\ND110Compile\uCode\CONT-STORE-10611.LISTING.TXT - Swapper disassembly: SINTRAN/ND500/swapper/swapper-k01-pseg.asm (+ raw .PSEG/.DSEG; NOTE: the disassembler DESYNCS after every loopi - it swallows the displacement byte into a long literal; the boundaries in this doc were re-verified from raw PSEG bytes where it matters, e.g. 1000077663) - String-dispatch carve: STRING-DISPATCH-BLOCK-CARVE-RELAY-2026-08-17.md - Protocol context: CARVE-ANSWER-N5SWAP-FUNCTION-VOCABULARY-2026-08-17.md, CARVE-ANSWER-N5SWAP-FN3-SLOT-TABLE-INIT-2026-08-17.md, N5SWAP-SWMSG-FIELD-DOSSIER-RELAY-2026-08-17.md (fn 5 = MSWIN, L07:5680) - Execution proof: SwapperN5Swap_Fn22B_SetConfig_ThroughJumpgTable (the four-function drive) + the new oracles in RetroCore/Nuget/HackerCorpLabs.Emulation.CPU.ND500UC/tests/, 191/191 green 2026-08-17. Engine model: package docs/TICK-MODEL.md milestone-14 section (INFERRED items 83-88).

1. The CPGU/CWIP exit anomaly RESOLVED (PROVEN words, INFERRED semantics - item 83)

The listing keeps TWIN dispatch terminators apart on bit 142 (SPAREBIT):

  • 002325B SPAREBIT JMPMAP PRF,EOP - the OUT-OF-LINE exit. Users: the page-op exit 010613B (A,P D,DP PRF,CLEAR -> 002327B PRF,START at DP = P), the float exits 002253B/002255B/002270B/002272B, the trap-path twin 002664B-002666B, and LOOPI's not-taken arm (002323B POPRET -> 002255B). The re-decode presents an address INSIDE the just-completed instruction; the EOP steps OVER it to the completed instruction's byte end.
  • 002221B JMPMAP PRF,EOP (NO sparebit) - the in-place dispatcher (RET/ifkret 002546B; the 0-operand system ops compute DP := P + 2 themselves, 007530B).

This RETIRES the milestone-12 "PRF,CLEAR resyncs P" model (the two are mutually exclusive on ZPGU, and only the SPAREBIT one lets operand-less CPGU/CWIP - and the F/D float exits and LOOPI - terminate). Pinned by execution: CPGU;CWIP back to back complete with P advancing, and the BRANCH_loopi golden file went 30/0.

2. The fn-5 loop-head bypass and SWMSG intake (PROVEN, executed)

  • 1000101504 w comp2 [0x240B8],$5; if = go $25 - fn 5 skips ONLY the slot-table gate compare (like fn 3's arm at 1000101515) and reaches the 29-way jumpg. Executed: fn echo [0x240B8] = 5, stub 1000101554, worker 1000077534.
  • RIOM count for fn 5 = 70 halfwords - exactly enough to include the message's TRAILING POINTER (hw 68-69, section 4).

3. The worker 1000077534's first-time path (PROVEN, executed)

  • 1000077551 w test [0o246374]; if = go $104: the init-done flag [0x14CFC] ships 0 in the DSEG -> block A (1000077561-1000077662: the re-init flush with hconv/47365 calls and a loopi'd release loop ending in ret) is SKIPPED; the first-time spine resumes at 1000077663.
  • 1000077663 w bmove [0x08030F8C],[0x0801285C],$3 + 1000077700 w move - config copies (raw-byte verified; the disassembler is desynced here).
  • Sub 1000110007 (whole, executed): w sub3 r.10,r.4,b.34 (the 176156B byte anchor), w2 mulad $4,r.0 (251B), a w4 sfill whose descriptor spans the STATS BLOCK - cells [0o461114]..[0o507454] - so the fn-5 init ZEROES the per-code stats INCLUDING the total-message counter [0x26280], then three 3-word w bmoves.
  • The jumpg helper 1000077515: entd; l=: b.100; ced=: [0o246454]; clrk; jumpg b.100 - executed; ced=: stores the CURRENT EXECUTING DOMAIN = the AM#15 main-domain register (0 here; the pairing is CONFIRMED by 76/76 MOVE_ced=_ golden vectors in the engine sweep).
  • Sub 1000065427 + callee 1000003444: stamps a descriptor into data segment 15B (VA 0x68000000+; by rladdr $15000000000+ with slot = [0o246454-family]*0x400; putbf/putbi at r3.(0)). Executed against a zero-mapped segment (staging, section 6).
  • The slot-table RESET: the fn-3 gate cell [0x128E4] goes 0x40 -> 0 (executed observable) - fn 5 re-initializes the slot table fn 3 built.
  • Sub 1000103415 (the DESCRIPTOR-BLOCK UNPACK, executed): takes the SWMSG hw-68/69 word = a PHYSICAL address (the spine reads it via r := b.24; w1 := r.210 = intake+136) and, under dmof (data mapping OFF - AL#17 bit 23B clear via 010547B, so the dereferences are PHYSICAL), unpacks: [0o461020] := h[block+2] (walk bound), [0o461024] := block+36 (list-head cell), [0o461030] := block+40, [0o461034] := block+0.
  • Sub 1000077433: reads the OPTIONS word (SWMSG hw 8) via [W1+0x10], h shl r2,$72, and with options = 0 SKIPS its 1000003057 announce (executed).
  • Sub 1000066175: derives the mode flags [0o436530..0o436570] from the options word (by2 := r1.(21) and $7, w getbi r1.(20),$23/$25), then calls 1000107011 (executed).
  • Sub 1000107011 (the MAILBOX LIST WALK, executed): under dmof, walks the node list at [[block+36]] - per node: done-flag h[node+8] (compared to 1), next = [node+4] - bounded by [0o461020]. Its TWO 1000003057 announce sites are ERROR arms (codes 0o200 = bound exceeded, 0o201 = no node seen), both SKIPPED with a healthy one-node list.
  • The 255-entry TABLE-B stamp loop (executed): 1000077731-1000077760 h set1 b.30; h wconv; w1*$400; by rladdr $4000004400+; h2 =: r.330; loopi b.30,$376 - stamps a halfword into every entry of data segment 4 (base 0x900, stride 0x400).

4. The fn-5 SWMSG (carved + executed geometry)

halfword consumer meaning
8 (+0o20) 1000077433 / 1000066175 OPTIONS word (low 3 bits + getbi bits 0o23/0o25 -> mode flags; 0 = the minimal path, no MON arms)
51 (+0o146) spine 1000100126 (AFTER the deferred wall) segment count -> [0o224124]
53/54/57 shared copier 1000064476 (record base intake+0o150; AFTER the wall) the three config cells [0x14D1C]/[0x14D20]/[0x1284C]
68-69 (+0o210) spine -> 1000103415 PHYSICAL address of the descriptor block (bound at +2, mailbox list head cell at +36) - why the RIOM count is exactly 70

5. The 1000003057 announce = SWPFATAL, not "sub-fn 2" (PROVEN bytes - CORRECTION)

The generic announce worker 1000003057's MON 377B (at 1000003167, argc 2) carries arg-1 = the CELL [0o225040] = [0x12A20], which ships 0o2047 = SWPFA (SWPFATAL) in the DSEG - the deep-analysis dossier's own site table lists it as the 0o2047 argc-2 site. The swapper-k01-handlers.md reading "MON 377B sub-function 2" is WRONG; the hot flags [0o436554]/[0o436560] gate the CACHE-FLUSH block (1000003123-1000003163), not the MON. The announce loop (1000003207 ifkret, 1000003210 go -0o21) re-issues until K = 1, and ifkret PROPAGATES K up every caller (call X; ifkret chains) - so a K = 1 reply unwinds the worker into 1000101241 (which writes W1 into the fn cell [0x240B0], clrk, returns to the stub). The codes the fn-5 error arms pass (0o110, 0o200, 0o201, 0o66 for the older handlers) are ERROR/EVENT numbers in PAR2. OPEN: the ND-100 side of this protocol (what SINTRAN does with SWPFU = 0o2047 beyond ESWPFATAL/XRSTARTALL, and whether any reply flavor lets the announce loop terminate NON-fatally).

6. Harness staging (INFERRED - TICK-MODEL item 87) and the deferred wall

Staged (all documented in SwapperAttemptTests): - Shared data segments as writable zeros: segment 4 (Table B, 132 pages), segment 5 (offset window 0x10000-0x20000 - sub 1000066411's $5000204000+ putbf stamps), segment 15B (96 pages). - The SWMSG hw-68/69 descriptor block + ONE healthy done-flagged mailbox node in physical 0xF8000 (bound 8, head -> node, node.flag=1, node.next=0). - [0o224014] = [0x1280C] (a slot pointer 1000066411 page-numbers; ships the stale live value 0x080B8000) -> the item-82 zero scratch.

THE WALL (deferred): past the mailbox walk, sub 1000071273's per-segment Table-A rebuild dives into SINTRAN-maintained descriptor state (frame chain: spine -> 1000071273 -> 67xxx -> the 2121-region paging sub -> a fatal announce whose own argument pointers are null) - page fault 46B during the announce's argument walk. Reaching cwip (1000100120), the copier re-run and the final set1 [0x14CFC] needs the uncarved SINTRAN Table-A/descriptor init. The attempt pins (test asserts): dispatch through the bypass arm, the execution footprint (SUB3/MULAD/SFILL/BMOVE/LOOPI>=254/SHL/GETBI entries), counter zeroed, gate reset, the unpack cells, and the recorded trap park (N5STA=3, STOPR=2, TRAPN=46B) with a clean return to IDLE.

7. What this closes and what stays open

Closed (PROVEN, executed): the fn-5 bypass arm; the worker's first-time spine through nine subsystems; the SWMSG hw-8/68-69 geometry; the descriptor-block unpack and mailbox-walk protocol shapes; the 1000003057 = SWPFATAL correction; the SPAREBIT exit (item 83); the instruction batch (section 3 families) with oracles.

Open: the Table-A/descriptor SINTRAN init (blocks fn-5 completion); the ND-100 reply protocol for the SWPFATAL announces; SWMSG hw 51/53/54/57 semantics past the wall (carved statically, unexecuted in fn 5); D MULAD / F-D LOOPI / LOOPD / SHA / SOLO entries (unmapped, unexecuted).

ADDENDUM 2026-08-17 (milestone 15): the ND-100 reply protocol RESOLVED

The section-5 OPEN ("what SINTRAN does with SWPFU = 0o2047, and whether any reply flavor lets the announce loop terminate non-fatally") is CLOSED, carved from the NPL nucleus (MP-P2-N500.NPL, CC-P2-N500.NPL, XC-P2-N500.NPL, RP-P2-N500.NPL) plus the 24B microcode path:

  • KFLIP bit 0 is the ONLY source of the resumed process's K flag (PROVEN both sides: the 3MONCO resume microwords 010212-010306 mask S1 bit 8 out and OR it back from KFLIP bit 0; OKMONICO sets T:=0, EMONICO sets T:=1 before the shared MONICO tail stores KFLIP := T, CC-P2-N500.NPL:359-366).
  • Legal announces get replied: new work via 5ACTSWAPPER (KFLIP=0, NUMPA=6 - writes the fn code into [0x240B0] and the requester's physical address into [0x240B4], MP:2884); transfer completions via OKMONICO (KFLIP=0) or EMONICO with code 0o1055 SWDERR (KFLIP=1 -> K=1 -> the swapper's ifkret chain unwinds = how it learns a disc transfer failed).
  • A SWPFA (0o2047) announce gets NO reply at all (PROVEN, MP:1192-1193): SWPFU > SWFMAX(6) -> ESWPFATAL -> XRSTARTALL, which errors every queued ND-500 message, restarts the ND-100 owner RT programs, sets FERROR := 0o2047 (N500TMR then stops watchdog servicing, RP:127653) and never re-activates the swapper process. The ND-500 side stays parked at the MON 377B stop forever.
  • Harness rule: a mid-worker SWPFA announce is a terminal park - never auto-resume. (A KFLIP=1 resume exercises the unwind mechanically but is NOT SINTRAN behavior for SWPFA; KFLIP=0 makes the announce loop re-issue forever.)

ADDENDUM 2026-08-17 (milestone 15): fn 5 driven TO COMPLETION

The section-6 wall is CLOSED. The completion carve (spine 1000077767 -> 1000100644, sub 1000071273's full body, the 1422/1755/1165 PTE machinery, 1000066411's two-level table constructor, 70054/75410/76133/ 65112/65622/72767) lives in FN5-COMPLETION-STAGING-CARVE-RELAY-2026-08-17.md (same folder), and the executable staging contract is in the engine harness (RetroCore/Nuget/HackerCorpLabs.Emulation.CPU.ND500UC/tests/ SwapperAttemptTests.cs, the fn == 5 branch) with TICK-MODEL item 89. Key corrections to THIS doc from execution:

  • The per-segment rebuild loop bound of 1000071273 is SWMSG hw 44 (intake+0o130), NOT hw 51; hw 51 is stored to [0x12854] by the spine AFTER cwip. A ZERO hw 44 is not a live shape: [0x128D4] := hw44 - 1 goes 0xFFFF and 1000076133's scan compare (w comp2 b.110,r1; if >> go) is UNSIGNED - the microcode loops 65535 times and walks off the DSEG.
  • The milestone-14 "trap 46B in the announce argument walk" reading is refined: the wall was 1000066411's very first descriptor write (1755 state-0/page!=0 -> announce 0o20 at 1000002121) with the announce faulting on unmapped DSEG pages; with real slot descriptors staged the announce never fires.
  • There is NO completion MON on the success path: the worker ends set1 [0x14CFC]; ret and the stub returns the swapper to its main-loop LNEWSWAP announce.

Completion post-script (same day, after the run went green)

The completion is EXECUTED: 195/195 package tests green with fn 5 running set1 [0x14CFC]; ret and returning to the main-loop announce. Execution corrected three more static readings (details in the relay doc and the engine's TICK-MODEL milestone-15 section):

  • The record TYPE-REMAP table at DSEG 0x24014 rewrites every raw type to 0 or 3 - the 4..7 per-code arm of 71273's loop #2 is UNREACHABLE with the shipped table. Raw type 4 (-> 3) is the arm whose tail writes the [0o224370] free-chain head; raw type 0 routes the head elsewhere.
  • The 71273 record tail BUILDS the free chain out of seg-6 entry indexes 1..count-1 and links its last entry to a frame-LOCAL zero, ORPHANING anything pre-staged at [0o224370] - the free chain IS the swap-in record's own page set. 72767's phase-2 bound b.64 divides the Table-A HEADER capacity word ([0o224014]+8), not a seg-6 cell.
  • Three ENGINE defects had to fall before the heap survived the four 76133 passes: the SHC register is 6-bit two's complement (else shl $66/$72 and sha $76 read +54/+58/+62 and the 1422/1755/1165 walk collapses every directory index to 0), EX,SHA (EXFUNC 10B) existed only half-wired, and EX,SHR rotates LEFT for positive counts. The swapper's own shipped code was right all along.