thumb2c tune: learn the HImode literal-pool respellings

[wowwheaties]

Summary

thumb2c tune learns the two HImode literal-pool respellings, so drafts whose only residue is an early literal pool can reach exact without hand editing. One site at a time, kept only on a strict differing_halfwords improvement, and only while the aligned report still shows the candidate's pool behind the reference's. Ordinary C89 source alternatives; no flags, no register or scheduling forcing; nothing outside tools/thumb2c/.

Mechanism (read from the compiler)

GCC 2.96 places a Thumb literal pool by the tightest pending fix, and each fix takes its range from the pattern alternative recog chose, not from the printed text (push_minipool_fix, arm.c:5367; entries sorted by max_address in add_minipool_forward_ref, arm.c:4829; the barrier chosen in create_fix_barrier, arm.c:5240). A symbol address spelled through SImode files a 1020-byte fix (*thumb_movsi_insn alt 6, arm.md:3838). Held in HImode it files a 64-byte one (*thumb_movhi_insn alt 1, arm.md:4318), which is where the ROM's pool sits in these functions. A plain u16 local does not help: PROMOTE_MODE (arm.h:618) widens it back to SImode.

• Construct A, a symbol-valued local: s32 p8; p8 = (s32)Sym; (or the one-line s32 p8 = (s32)Sym;) → u16 p8[1]; p8[0] = (u16)(s32)Sym; with every read respelled p8[0]. Refused unless the local is declared once, assigned once from a symbol the unit declares extern, and read only into halfword-or-narrower stores.
• Construct A′, a symbol value stored straight into a halfword-or-narrower target with no local (rec7[34] = (s32)Sym;): a one-element u16 temporary is declared beside the other locals and the store goes through it.
• Construct B, a halfword store of a symbol-derived constant: *p = (u16)(s32)Data_00000e00; → *p = 0xe00; (the lifter names such a word by the value the image holds; only values in 1..0xffff qualify).
• Guard pool_is_late(report): an aligned row where both columns load a pool word and the candidate's is at the higher address.

Fixtures (overlay score <file> --owner <owner> --span <span>)

Owner	span	draft	hand-tuned	thumb2c tune from the draft
resource_3c9:0200423c	1692	114	0	0 (114 -> 0 himode p8[1])
resource_3a4:02001838	1236	531	21	20 (531 -> 111 himode p10[1], then two call respellings)
resource_371:020039fc	1296	513	66	98 (five himode *shown = 0x…; steps; the guard turns false after the fifth site, so it stops short of the hand file by design)
resource_3ba:02002aec	192	69	1	69 → 45 on the first step (himode data[1], the one-line initializer form); the rest of the hand file is a shared zero and store order
resource_3bb:02002d84	192	67	10	67 → 42 on the first step (himode store, the direct-store form); the residue is allocation-shaped

Sweep over the 20 remaining near-miss overlay drafts on 7d8a720 (thumb2c tune from each lane's best draft): resource_3af:02002618 312 → 43 in one himode store step, resource_3b9:02001298 346 → 218 and resource_3b7:02000e5c 838 → 729 through the existing call respellings; the other owners moved by 0–4.

The remaining residue on 3a4 and 371 is allocation and scheduling shaped and goes to the smart queue.

Gates

• cargo test (thumb2c): 16 passed; five new unit tests beside respells_calls_both_ways.
• make test, make lint (forbidden=0), make tooling-index-check (87 tools), make verify: green.
• make tooling-size: 46,996 / 50,000.

Review notes

• Two commits: the mechanism and guard, then the initializer and direct-store forms found by two more drafts. Built on c379b65; merges clean on 7d8a720. Default output is unchanged for units where the report never shows a late pool.
• The evidence write-up with the arm.c line references is in the branch's scratch/himode/REPORT.md (ignored path, not part of the commit) and was published earlier on ☀️ 73% – adopt seventy-eight exact owners on 2338f8f0a: ten main-image functions and sixty-eight scene sequences #10.

🤖 Generated with Claude Code