stefanheule
commented
8 years ago I've implemented this, and about 65.8% of the latencies stayed the same (though some of them are fixed to 0 or 999). For the rest, they mostly went down (because there fewer dependencies now between instructions executed one after the other). For some of the instructions with very high latency (say > 300, e.g. the x87 instructions), the numbers are almost the same, they just changed a little due to random fluctuations. So likely the true percentage of instructions without any significant change is higher.
Some noteworthy things:
- The non-temporal hint instructions got much more expensive (from 1 to ~70). Not sure why, but I don't think this matters.
- All cmov* instructions are now cost 1 (memory or not).
See https://github.com/StanfordPL/stoke/issues/785