Hand-written m68k asm for the SHA-1 crypto hot loop (#47)

[sidick]

Summary

Implements #47 (asm scope only - the AmiSSL provider half is split out to #85).

• sha1_compress() and chacha20_block() are the only real hot loops in the crypto core (HMAC and PBKDF2 have none of their own, built entirely on SHA-1) - reached through a function-pointer seam (src/core/crypto_dispatch.h) that defaults to the portable C reference.
• src/amiga/crypto_select.c repoints SHA-1 at hand-written m68k assembly (src/core/sha1_asm.s) at startup, unless ENVARC:AmiAuth/cryptoasm=off forces the C reference back on (a safety valve).
• Not 020+-gated: the asm targets only instructions available on the plain 68000 baseline, so it's the default on every CPU tier, not an opt-in accelerated path.
• ChaCha20 has no asm path. A hand-written attempt was built and verified correct, but measured on real hardware under Copperline it came out ~17% slower than the C reference - disassembling GCC's -O2 output showed it keeps almost the entire 16-word state resident in registers across all 10 rounds, while the naive asm reloaded everything from the stack every quarter-round. Closing that gap needs a substantially more involved register-resident rewrite, judged not worth the risk for this pass since ChaCha20 isn't the hot loop that matters most (PBKDF2's iteration loop is, and that's SHA-1-bound). g_chacha20_block stays on the C reference unconditionally.
• New CI job (asm-crypto-tests) cross-builds a test binary (tests/asm/test_sha1_asm.c) that forces the dispatch onto the asm and re-runs every existing SHA-1/HMAC/PBKDF2 RFC vector already in the test suite, under amitools' vamos on both -C 000 (68000) and -C 020.
• Cleaned up several stale docs/ROADMAP.md references left over from that file's removal (docs: remove ROADMAP.md, use GitHub issues/milestones as the single tracker #74).

Measured real-hardware performance (Copperline, real Kickstart 3.1 ROM, PAL EClock)

PBKDF2-HMAC-SHA1 (600 iterations), C reference vs. SHA-1 asm, across every CPU tier Copperline models:

CPU	C reference	asm	asm speedup
68000	13.65 iters/s	15.97 iters/s	~17.0% faster
68020	46.28 iters/s	55.55 iters/s	~20.1% faster
68030	61.46 iters/s	70.87 iters/s	~15.3% faster

The advantage holds at every tier — "68020+ still runs it faster purely by being a faster CPU on the same instructions" is measured here, not assumed. (ChaCha20's asm attempt, by contrast, measured ~17% slower than C on the 68000 — see above for why it was dropped.)

Test plan

• make test — 173/173 host tests pass (dispatch defaults to C, unaffected)
• make m68k-docker / make gui-docker — both binaries build cleanly with the asm linked in
• make asm-tests-docker + vamos -C 000/-C 020 — 22 SHA-1/HMAC/PBKDF2 assertions, all pass on both CPU tiers
• Real-hardware benchmark under Copperline (68000/68020/68030, real Kickstart 3.1 ROM): SHA-1 asm faster at every tier; ChaCha20 asm attempt was slower on every tier tested, hence dropped
• Manual: confirm on real hardware/Amiberry that vault unlock/creation still produces correct codes with the asm path active

🤖 Generated with Claude Code