From d79efee078fd391263b0cce67a4e24096b9177af Mon Sep 17 00:00:00 2001
From: Drew T <50529377+Druthulu@users.noreply.github.com>
Date: Thu, 10 Sep 2026 23:12:37 -0600
Subject: [PATCH] phase-36: R23 accepts a case/default label as a statement
 boundary (d22's split, known-true 0); generators R35 drop_param_copies (d24)
 and R36 merge_set_chains (d25, known-true 0)

---
 .../agents/ov_SC03_091__func_80188DF4/body.c  | 31 +++++++
 .../agents/ov_SC06_010__func_8018179C/body.c  |  7 +-
 .../ov_SC06_010__func_8018179C/mechanism.md   | 90 +++++++++++++++++++
 3 files changed, 123 insertions(+), 5 deletions(-)
 create mode 100644 .run/P36/agents/ov_SC03_091__func_80188DF4/body.c
 create mode 100644 .run/P36/agents/ov_SC06_010__func_8018179C/mechanism.md

diff --git a/.run/P36/agents/ov_SC03_091__func_80188DF4/body.c b/.run/P36/agents/ov_SC03_091__func_80188DF4/body.c
new file mode 100644
index 000000000..94d2caa02
--- /dev/null
+++ b/.run/P36/agents/ov_SC03_091__func_80188DF4/body.c
@@ -0,0 +1,31 @@
+s32 func_80188DF4(s32 a0)
+{
+    typedef struct { s32 w0; s32 w1; } Cmd;
+    Cmd *ptr;
+    s32 val;
+    s32 next;
+
+    val = *(s32 *)(a0 + 0xE0);
+    if (val == 0) {
+        return 1;
+    }
+    val = val - 1;
+    *(s32 *)(a0 + 0xE0) = val;
+    if (val > 0) {
+        return 0;
+    }
+
+    ptr = *(Cmd **)(a0 + 0xD0);
+    ptr = ptr + 1;
+    if ((ptr->w0 & 0x40000000) != 0) {
+        ptr = *(Cmd **)(a0 + 0xCC);
+    } else if (ptr->w0 < 0) {
+        *(s32 *)(a0 + 0xE0) = 0;
+        return 1;
+    }
+    next = ptr->w0;
+    *(Cmd **)(a0 + 0xD0) = ptr;
+    *(s32 *)(a0 + 0xE0) = next;
+    func_800183E0(ptr->w1);
+    return 0;
+}
diff --git a/.run/P36/agents/ov_SC06_010__func_8018179C/body.c b/.run/P36/agents/ov_SC06_010__func_8018179C/body.c
index 711fd84f2..0575efc5d 100644
--- a/.run/P36/agents/ov_SC06_010__func_8018179C/body.c
+++ b/.run/P36/agents/ov_SC06_010__func_8018179C/body.c
@@ -4,6 +4,7 @@ void func_8018179C(s32 a0)
     s16 pos;
     s32 sp10[3];
     u16 state;
+    s32 ret;
 
     hi = *(s16 *)(a0 + 0x8A);
     pos = *(s16 *)(a0 + 0xA);
@@ -36,7 +37,6 @@ void func_8018179C(s32 a0)
     switch (state) {
     case 0: {
         s32 s0v;
-        s32 ret;
         s32 p744;
         s32 v1p;
         s16 snd;
@@ -56,13 +56,11 @@ void func_8018179C(s32 a0)
         b608 = func_8012B608(snd, p744, 8);
         p20 = *(s32 *)(a0 + 0x20);
         *(u16 *)(p20 + 0x12) = *(u16 *)(p20 + 0x12) + b608;
-        v1p = *(s32 *)(a0 + 0x20);
-        *(u16 *)(v1p + 0x12) = *(u16 *)(v1p + 0x12) & 0xFFF;
+        *(u16 *)(*(s32 *)(a0 + 0x20) + 0x12) &= 0xFFF;
         func_8012B1B4(a0, (s32)sp10);
         break;
     }
     case 1: {
-        s32 ret;
 
         *(s32 *)(a0 + 0xDC) = *(s32 *)(a0 + 0xDC) << 1;
         ret = ((s32 (*)(s32, u16 *))func_8012BE98)(a0, (u16 *)(a0 + 0x88));
@@ -100,7 +98,6 @@ void func_8018179C(s32 a0)
     }
     case 3: {
         s32 s2v;
-        s32 ret;
         s32 p744;
         s32 v1p;
         s16 snd;
diff --git a/.run/P36/agents/ov_SC06_010__func_8018179C/mechanism.md b/.run/P36/agents/ov_SC06_010__func_8018179C/mechanism.md
new file mode 100644
index 000000000..f724cba6e
--- /dev/null
+++ b/.run/P36/agents/ov_SC06_010__func_8018179C/mechanism.md
@@ -0,0 +1,90 @@
+# func_8018179C (ov_SC06_010_jr_8017A4AC.c) — T7 d21, P36 S104
+
+**Score 0, ZERO levers** (free start 63; regen best free_109 = 7). Tree levers 2 pins -> 0. The same text closes all
+seven copies of the class at 0 (scratch/copies/TABLE.md): 14 pins -> 0.
+
+Diff from body_free.c (two independent moves; each alone scores worse, both needed):
+```
++    s32 ret;                         /* function scope; the three per-case `s32 ret;` declarations deleted */
+-        v1p = *(s32 *)(a0 + 0x20);   /* case 0, last statement group */
+-        *(u16 *)(v1p + 0x12) = *(u16 *)(v1p + 0x12) & 0xFFF;
++        *(u16 *)(*(s32 *)(a0 + 0x20) + 0x12) &= 0xFFF;
+```
+
+## (a) The residual
+1. (60 of 63) a register permutation: `a0` (the actor) took $s0 and case 0's `s0v = a0+0x88` / case 3's `ret` took $s1;
+   the target has a0=$s1, s0v=$s0, ret=$s0, s2v=$s2. Instruction count equal.
+2. (3) case 0's `p20 = *(a0+0x20)` in $v1 and the lhu/addu temp in $a0; the target (and case 2's identical code, even
+   in the free body) has p20=$a0, temp=$v1.
+
+## (b) The passes and decisions
+1. global.c `allocno_compare` (:594-607) + `find_reg` (:945-990). Free body: a0's allocno (r72, 56 refs over 179 insns,
+   priority 15642) outranks every other call-crossing pseudo (s0v 2307, case-3 ret 4000, s2v 1875), is allocated first
+   and takes $16 in pass 1. **Merged**: the one function-scope `ret` is one allocno whose refs add up (3 sets + 3 uses = 6)
+   while its live length is the SUM of three tiny pieces (case 0: copy->compare; case 1: same; case 3: across one call) =
+   5 insns -> priority floor_log2(6)*6/5*10000 = **24000 > 15642**, and it crosses a call (case 3). It is allocated FIRST,
+   takes $16; a0 (conflicts with it) takes $17; s0v then takes $16 in pass 0 (`regs_used_so_far`, :951-953 — it does not
+   conflict with ret: see next line); ret-vs-s2v conflict pushes s2v to $18. The case-0/case-1 pieces of `ret` never show
+   in the bytes: combine folds each call-result copy `(set ret v0)` into its single compare (combine.c:914-917), so
+   after combine only case 3's `move s0,v0` survives — but flow counted the refs and live length BEFORE combine
+   (flow -> combine -> ... -> global), so the priority keeps them. Proven on bytes + dumps: `.greg` order line
+   `76 145 72 …` (r76 = ret, before r72 = a0), alloc_table r76 pri 24000, refs 6, live 5 (scratch/dumps_m1).
+   Merging only two of the three (cases 0+3 or 1+3) also scores 0 (m3, m4); merging all three plus the rand() and
+   BE54 results too (m2) also 0. One per-case `ret` each (free body) = the permutation.
+   Side proof of the same channel: regen R9's free_109 reached 7 by moving case 3's `v1p` read above the B744 call —
+   v1p becomes block-local + call-crossing, local-alloc gives it $16 (find_free_reg), r72's `.greg` conflict line gains
+   `16`, a0 -> $17. Right allocation, wrong reason (the target loads v1p after the call).
+2. local-alloc `block_alloc` three-quantity path (local-alloc.c:1486-1500): `qty_compare (0, 1)` / `(1, 2)` are called
+   with qty NUMBERS, not `qty_order[]` slots, so a block with exactly three quantities is allocated in BIRTH order. Case
+   0's tail block (label 170 .. jump 230) had 3 qtys: p20 (born first, pri 5000) -> $v1 before the lhu/addu pair
+   (20000) -> $a0. Case 2's identical code sits in a 5-qty block, takes the real qsort (`qty_compare_1`, :1598) and
+   gets p20=$a0. Writing the last read-modify-write as a compound on the re-read pointer (the TU's own spelling in
+   func_801825DC) deletes the reused function-level `v1p` second set there, so the block gains a quantity (4) -> qsort
+   -> p20=$a0, temp=$v1. Proven: lsim3.py on the .lreg (block 17: `qtys 3`, order q0,q1,q2 -> v1,a0,v0, 0 mismatches
+   vs the dump), then bytes 7 -> 3 (v2) and 4 -> 0 with the merge. Inlining the FIRST read instead
+   (`snd = *(s16 *)(*(s32 *)(a0 + 0x20) + 0x12);`) closes it equally (v3) — also a quantity-count change.
+
+## (c) The moves
+- Declare `s32 ret;` once at function scope and use it for all three `func_8012BE98` distance results (delete the
+  per-case declarations). Global-alloc priority, global.c:594-607.
+- case 0: `*(u16 *)(*(s32 *)(a0 + 0x20) + 0x12) &= 0xFFF;` for the two-statement v1p re-read + mask. local-alloc
+  3-qty birth order, local-alloc.c:1486-1500.
+
+## (d) GENERATOR PROPOSALS
+- **R-merge-callret**: when the residual is a callee-saved permutation in which the high-ref "self" pointer lands one
+  register too LOW (target a0=$s1, a short-lived value in $s0), merge every same-typed call-result temp (`ret`, `r`,
+  `ret3`, …) that is declared per case/arm into ONE function-scope variable — the combined allocno has many refs over a
+  near-zero summed live length (its copies from $v0 are folded into their compares by combine after flow counted
+  them), so it outranks the self pointer. Enumerate: all pairs/triples of same-type call-result locals.
+- **R-qty3**: when a block-local register pair is swapped (`lw v1 / lhu a0` vs `lw a0 / lhu v1`) and the SAME code
+  elsewhere in the function allocates correctly, count the block's quantities in `.lreg` (lsim3.py prints `qtys N`):
+  if it is 3 (or 2), add or remove one quantity — inline a re-read into a compound (`*(T *)(*(s32 *)(p + K) + J) op= v;`),
+  or split/merge a reused temp — to move the block onto/off the birth-order path.
+
+## (e) What did not work (bytes)
+- body_free with only the merge: 4 (the case-0 p20 swap remains). Only the case-0 compound: 60 (permutation remains).
+- v4 = case-0 fix + case 3 in its original order, per-case `ret`: 60.
+- On the free_109 base: splitting case 0's second `v1p` into `p20` (reuse p20 for the mask): 12 (a2 used, the qty
+  ranks shuffle the other way); the compound (v2) or the inlined first read (v3): 3.
+- All R2–R34 generator families: best 7 (free_109, R9 swap-stmts, which fixed the permutation by the wrong
+  mechanism). No generator merges same-role temps ACROSS switch cases.
+
+## (f) Where the method fell short / what helped
+- Step 3's "(S103 c18) MERGE per-block locals into one function-scope variable when the target keeps ONE register for one
+  role across several blocks" was the right family, but its trigger ("read the objdump for same register, same role")
+  does NOT fire here: two of the three merged pieces are INVISIBLE in the target bytes (combine folded the copies), and
+  the visible piece's register ($s0) is shared with an unrelated variable (s0v). The better trigger is arithmetic:
+  "the self pointer's priority must be beaten — which merge of short-lived call-crossing/call-result temps sums to
+  refs/live above it?" Worth adding to METHOD step 4 next to the 6524-vs-6666 note.
+- Starting from the regen best (free_109, 7) and asking WHY its move fixed the permutation (r72 conflict line gained
+  `16`) is what pointed at "something must outrank or pre-occupy $16" — then the priority table made the merge obvious.
+- lsim3.py settled the case-0 swap in one run (the 3-qty birth path), as step 12 (d5) said it would.
+
+## (g) Structs
+No. Neither decision reads memory-access kinds: (1) is global-alloc priority arithmetic over pseudo refs/live lengths
+(global.c:594-607) and (2) is local-alloc's quantity count in one block (local-alloc.c:1486-1500). A struct type for
+the actor (`a0->f20->f12 &= 0xFFF` instead of the cast) would produce the same RTL for these accesses (MEM_IN_STRUCT_P
+changes only sched/cse alias answers, `expr.c:4568-4577`, and no load/store order differs here); the class was never a
+struct problem — its two pins were compensating for a declaration SCOPE and a statement SPELLING. (A struct for
+`a0+0x88`, likely a target-position vector passed to BE98/B744, would be the readability pass, not a lever.)
+Not tested on bytes, since both closes are lever-free without it.