* timing: use single ported SRAMs, invalidating read responses on write
* performance:
-- shortening history length to accelerate training
-- use a predictor to reduce s2_redirects on FTB not hit