The second pass concluded the 168 B tricks->asm gap was per-call ABI
cost. Most of it was structure. Rebuilt around the oracle's own shape —
argless noinline primitives over a whole-loader call-saved register
protocol (g_addr in Y, count r16, window r7, direction latch r6), a
top-down erase_below whose loop tests against zero and hands callers
g_addr = 0 for free, bounded rx everywhere (a silent host unwinds to
the app from any state, as the oracle does), and a named tsb_app entry
that --pmem-wrap-around=32k relaxes to the wrapped rjmp:
tsb_asm 510 B in the 512 B section (oracle: 500), C++ except rx
and the page-store loop — the two routines whose remaining
cost is the calling convention itself (~30 asm lines, was
~280)
tsb_tricks 526 B, no assembly at all (was 666)
tsb_pure 836 B, still one readable function per command (was 842)
Every g_* update placement works around a GCC 16.1 wrong-code bug
(stores into global register variables deleted when only callees read
them — repro and rules in libavr dev/lessons.md). Also fixes two
latent hardware bugs all earlier tiers carried, masked by simavr's
zeroed register file: the crt-less entries never established
__zero_reg__ = 0, and the direction latch was read before written —
power-on registers are undefined.
All tiers full oracle feature parity, protocol tests green in both
libavr modes, .text byte-identical across modes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
BlackMark
74d8b92885