fix: align all 3 secp256k1 benchmarks for fairness and comparability
JVM benchmark fixes:
- signSchnorr(cached pk): document that native baseline still derives
pubkey (apples-to-oranges by design — shows what caching enables)
- privKeyTweakAdd: document .copyOf() penalty on native side (ACINQ
wrapper mutates input, requiring defensive copy)
- Remove duplicate tweakMulCompact test (redundant with ecdhXOnly)
- Add cache-warming note to benchmark KDoc
- Change output "slower" → neutral "x" (0.7x isn't "slower")
Android benchmark fixes:
- Remove misnamed pubKeyCompressOurs (was identical to compressedPubKeyForOurs)
- Remove duplicate pubKeyCompress (was also doing create+compress)
- Clean up structure: matched Native/Ours pairs with clear section headers
- Add cache-warming note to benchmark KDoc
- Standardize test data comment for cross-platform comparability
K/Native benchmark fixes:
- Fix DCE risk in micro-benchmarks: use result-dependent chains
(out→out for field ops, xor hiSink for multiplyHigh) so LLVM can't
hoist constant computations out of the loop
- Add batch verify (was missing — JVM had it, Android/K/N didn't)
- Add -opt documentation warning in KDoc
- Remove unused pubkeyCreate/pubKeyCompress standalone benchmarks
(compressedPubKeyFor already covers create+compress)
- Extract printResults helper to reduce duplication
All 3 benchmarks now:
- Use the same test data (same hex keys across JVM/Android/K/N)
- Measure the same core operations (verify, sign, compressedPubKeyFor,
secKeyVerify, privKeyTweakAdd, ecdhXOnly/pubKeyTweakMulCompact)
- Document cache-warming effects in their KDoc
- Include batch verify (JVM + K/N; Android uses framework-managed tests)
https://claude.ai/code/session_01EMY5RnXb9rnsyU2KbXrSaY
Claude
0c27ed89bc