[Repo Assist] perf: word-at-a-time backward copy in l_memmove

[github-actions]

🤖 This PR was created by Repo Assist, an automated AI assistant.

Summary

The backward-copy branch of l_memmove (triggered when dst > src and the ranges overlap) previously used a plain byte-at-a-time loop. This PR applies the same word-at-a-time technique already present in the forward branch.

Root cause

The forward path was already optimised (PR #107 era), but the backward path was left as:

while (len--)
    *--d = *--s;

For large overlapping copies (e.g. memmove(p+1, p, 64KB)) this is 8× slower than necessary on 64-bit targets.

Fix

Three-phase backward copy (mirrors the forward path):

1. Tail alignment — byte-copy trailing bytes until d is word-aligned.
2. Word-at-a-time — if s is also word-aligned, copy sizeof(uintptr_t) bytes per step backward using may_alias word pointers.
3. Head tail-out — byte-copy remaining leading bytes.

Read-before-write safety is preserved: in the backward pass d > s throughout, so each word write always lands at a higher address than the corresponding word read. No source data is overwritten before it is consumed.

Tests added (test_memmove)

Test	Purpose
Large backward overlap 64B (dst = src+1)	Exercises the new word-at-a-time path
Large forward overlap 64B (dst = src-1)	Confirms existing forward path handles large moves
Large non-overlapping copy 64B	Baseline sanity check

Test Status

Target	Build	Test
Linux gcc	✅ PASS	✅ PASS (1585 assertions)
Linux clang	✅ PASS	✅ PASS (1583 assertions)
ARM/AArch64	⚠️ SKIP (cross-compilers unavailable in CI env)

Generated by 🌈 Repo Assist at {run-started}. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@1f672aef974f4246124860fc532f82fe8a93a57e