Reading builtin80's body (0x82268460) to name it: it is NOT a predicate. It allocates a 20-byte object, stamps vtable 0x820A8CB0, magic 0xAB0311BA and the unit's live object into it, pushes it onto a queue via the same helper push.i uses, and returns 1 -- or 0 when the unit is absent. A command. That made the conditions listing impossible: it showed a six-way switch `if builtin80(TCT206) == 0 … == 5` on a function returning 1 or 0. Disassembling the site shows two unconditional `jmp`s between the call and the compare, so 0x1B6C0 is reached ONLY by a branch and its special[0] has nothing to do with builtin80. op12 is unconditional -- the next instruction is never reached by fall-through -- and the tracker walked through it exactly as it had walked through end_coroutine. Last iteration I fixed the instance and not the class, leaving 22x more bad sites in place than the fix removed. A/B over all 28 stages, 7563 sites, resetting at jmp as well: sites whose operands change 889 (11.75%) LHS unresolved, before -> after 34 (0.45%) -> 756 (10.00%) So the previous commit's headline "0.0% unresolved" was a MISSING CHECK, not a strong result: the linear walk always had some value to report, and reporting it was the bug. 10% is the honest figure and the other 90% is trustworthy for a reason. Also corrected: isl-unit-args.md illustrated its diff with 0x1B6C0, which is one of the bogus sites. The UNIT_ARG result itself stands -- it came from reading implementations, not from this listing -- but the example was picked from bad output. Not done, and said so: recovering the 756 needs a dataflow join over each block's actual predecessors, a CFG fixpoint rather than a linear pass. The branch targets are all known so the CFG is available; the analysis is not written. calls and phase-ends regenerate byte-identical; conditions changes on 187 lines.
4.3 KiB
✅ Which built-ins take a unit — read from the implementations, and the old set was short
isl.py's UNIT_ARG decides whether a built-in's slot-4 operand is printed as
a unit name or as a raw number. It was built statistically, from operand
ranges, and its own comment says slots were listed "only when the ratio stayed
below 1.0" — i.e. only when every observed value resolved. That is
conservative, and it was.
The direct method
isl-builtin-dispatch makes this a lookup: a
built-in's stub tail-calls a fixed slot of the ScriptPhase vtable at
0x820A84BC, so the implementation can simply be read. Every unit-taking
built-in opens the same way — this is unit_state, builtin80, builtin117
and builtin136, character for character:
lwz r10, 324(r31) ; the unit array
lwz r11, 4(r30) ; r30 = the local[] base, so this is local[4]
rlwinm r11, r11, 2, 0, 29 ; x4
lwzx r11, r11, r10 ; -> the record
lwz r11, 4(r11) ; the handle
cmpli cr6, 0, r11, 0x0 ; "is this squadron gone?"
✅ Result: 31 → 55
built-ins whose implementation indexes [phase+324] by an argument |
55 |
| of the statistical set's 31, confirmed | 31 — all of them |
| UNIT_ARG claims a unit, the implementation does not | 0 |
| implementation says unit, UNIT_ARG missed it | 24 |
The 24: 21, 22, 23, 32, 42, 44, 46 (squadron_trace), 49, 50, 51, 55, 60, 61, 72 (group_ratio_pct), 80, 83, 94 (is_engaged), 101, 109 (set_unit_flags), 117, 136, 137 (wait_units_ready), 141, 142 (deploy_and_wait).
Zero false positives is worth stating on its own: the statistical method was right about everything it claimed and only too cautious about what it omitted.
⚠️ The first control I chose was worthless — recorded because it nearly passed
I first checked whether the additions' slot-4 operands resolve to a symbol-table-2 index. They did, 100.0 % — and it means nothing:
| set | operands resolving to a symtab-2 unit |
|---|---|
| the 31 baseline | 100.0 % |
| the 24 additions | 100.0 % |
| the 92 built-ins in neither set | 99.3 % |
Symbol table 2 is dense enough that almost any small integer lands in it, so the test does not discriminate. A control that the negative class also passes is not evidence, and this one was one careless glance from being written up as proof.
✅ The control that does discriminate
isl-builtins.md documents that a symbol operand is a two-word pair — a tag
holding the constant 1, then the index. So slot 0 should be 1 exactly when
slot 4 is a unit:
| set | calls with slot0 == 1 |
|---|---|
| the 31 baseline | 100.0 % (13 677 calls) |
| the 24 additions | 100.0 % (140 calls) |
| the 92 in neither set | 2.5 % (2 903 calls) |
A 40× separation, and the additions sit exactly on the positive class.
Effect on the artefacts
data/isl-stage02.txt and -phase-ends.txt regenerate byte-identical;
-conditions.txt changes on 28 sites, every diff line pairing, all of them a
raw number becoming a unit name:
- 0x01B6C0 if builtin80(1, 117) == 0 -> + 0x01B6C0 if builtin80(TCT206) == 0
⚠️ That example was itself a bogus site — see
isl-conditions.
0x1B6C0 is reached only by a branch, so attributing its comparison to
builtin80 was wrong, and builtin80 is a command rather than a predicate. The
UNIT_ARG result is unaffected — it was derived by reading implementations, not
from this listing — but the illustration was picked from bad output.
🟡 Not settled
- The 24 are still unnamed. Knowing an argument is a unit is not knowing what
the built-in does.
builtin80(69 operands disc-wide) is tested against 0, 1, 2, 3, 4 in a switch chain, so it returns a small enumeration — but I am not naming it from that, and its body past the liveness check is unread. builtin103(0x8226BFA8) is a predicate over[phase+10152]and[phase+10156], neighbours of a value read out at+10160; it takes no unit. The fields have 9, 7 and 1 writers respectively, none of them read.builtin105(0x8226BFF0) tests a unit record's+16against 4. Whatrec+16holds is not established —isl-builtins.mdonly rules out its being whatunit_statereads.