| 1 | by Saam Barati, Michael Saboff, Tadeu Zagallo, Yusuke Suzuki, Robin Morisset |
| 2 | |
| 3 | == Now: |
| 4 | |
| 5 | - New Bytecode |
| 6 | - reduced bytecode size by 65% |
| 7 | - 12-13% less memory usage on a page like Facebook |
| 8 | - can be cached (not being done in Safari yet) |
| 9 | |
| 10 | - JetStream 2 |
| 11 | - join JS benchmarks together into a single suite (JetStream 1, ARES-6, RexBench, ...) |
| 12 | - 64 subtests emphasizing startup perf, peak throughput perf, worst-case perf |
| 13 | |
| 14 | - RAMification |
| 15 | - memory benchmark for JS |
| 16 | - leverages JetStream 2 tests, measures peak & current usage |
| 17 | |
| 18 | - ExecState -> GlobalObject refactor |
| 19 | - less error prone; most cases that were passing call frames didn't really need to be |
| 20 | - prereq for IsoSubspace work |
| 21 | |
| 22 | - WASM Interpreter |
| 23 | - improves startup time (3x), neutral on throughput |
| 24 | - perf is now similar to V8 / SM |
| 25 | |
| 26 | - Fuzz Zero |
| 27 | - improve security & stability by fuzzing |
| 28 | - fix all bugs and allow no regressions |
| 29 | |
| 30 | == 2020+: |
| 31 | |
| 32 | - BigInt |
| 33 | - ship |
| 34 | - JIT improvements |
| 35 | - new benchmark |
| 36 | |
| 37 | - Turbo DFG: |
| 38 | - vs. llint: baseline 2x, DFG 11x, FTL 17x |
| 39 | - DFG does not impact page load tests |
| 40 | - version of DFG that's less optimized but quicker to compile may help |
| 41 | |
| 42 | - Fast for-of: |
| 43 | - as fast as old-school for! |
| 44 | - involves inline caching for next iteration |
| 45 | |
| 46 | - IsoHeap Everything |
| 47 | - allocate like types together, prevent type confusion attacks |
| 48 | - do this for all JSObjects |
| 49 | - avoid memory/perf regression |
| 50 | |
| 51 | - Software Verified JIT |
| 52 | - constrain where calls/jumps go |
| 53 | - create a whitelist for far calls |
| 54 | - known near calls (e.g. for slow paths) can be whitelisted too |
| 55 | - pointer authentication |