The builtin Copy builds its result a char at a time (30x slow on ESP32)

Reported 2026-11-01 by the remote_thermostat session, measured on a classic ESP32 bridge board (xtensa, IDF 6.0.1). Parsing 33 mesh lines of ~470 chars took 1032 ms with Copy, and 37 ms after the app swapped in its own SetLength + Move. Mesh firmware-update throughput went from 0.5 KB/s to 7.2 KB/s, mostly because of that swap.

Cause. __pxxStrCopy in compiler/builtin/builtin.pas, which every program without sysutils gets for Copy, looped r := r + s[i]. That is one append, and so potentially one realloc, per char. Delete and Insert were built from two such Copies plus a concat. sysutils' own Copy already used SetLength + Move.

Fix. Copy is one SetLength plus one __pxxMove, and it re-reads Length(Result) so a frozen (short) Result is still clamped to its capacity. Delete and Insert each build the new string with one SetLength and up to three Moves. The same commit fixes sysutils' Copy, which formed index + count - 1 before capping count: Copy(s, i, MaxInt) overflowed and returned ''.

Measured on x86-64 under callgrind: 400 Copies of ~470 bytes went from 51.9M instructions to 0.6M (about 87x). The algorithm is the same on every target.

Test. test/test_copy_delete_insert_are_one_move.pas (+ .expected) covers every clamp edge, Copy(...)[k], a string[5] result, a shared source left untouched, and a 5000-byte string. Its output is identical to FPC 3.2.2. It runs in test-core on x86-64 and riscv32, and also matched by hand on xtensa, i386, aarch64 and arm32.

Confirmed on hardware 2026-11-01 by the thermostat session, ESP32 bridge board, d1a3dc081d built through tools/pxx_latest.sh, with the app's workaround removed. NextWord on 33 lines: 1032 ms (old builtin), 37 ms (workaround), 40 ms (new builtin). The whole S command went from 1354 ms to 215 ms, and a 32-frame round trip from 2.8 s to 0.69 s. No libgcc helpers.