3 ms·
Even with store-load fw, you get a penalty (~3 cycle latency) over register accesses, no?
by temac 7y ago
Even with store-load fw, you get a penalty (~3 cycle latency) over register accesses, no?
- Jasper_ 7y agoyeah, but it's cheaper than full L1 hit, which is where it would go if not for that.
- temac 7y agoI was trying to cite a typical full L1 hit latency... I thought store-load fw simply avoid having to flush the complete write buffer before the access is even possible, which risk to take far more than ~3 cycles. Now maybe it can be faster in some cases than an L1 hit, I don't know. Edit: it seems that store-load forwarding is actually slightly slower than L1: https://www.agner.org/optimize/blog/read.php?i=854#854 https://www.agner.org/optimize/blog/read.php?i=854#854