Aceshardware

Gabriele Svelto · Joined: 27 Jun 2007 Posts: 274 Location: Milano, Italy

Paul DeMone wrote:

The Alpha EV6 handled load speculation in a quite efficient fashion.
Load hit speculation is a latency win compared to a non-speculative
stall when you hit L1 but a latency loss if a miss occurs and a replay
is required. The EV6 included a miss predictor that kept track of the
hit/miss history of loads in a similar manner to branch predictors.
Loads that missed a lot (e.g. streaming data) were not speculated
and average performance was improved by avoiding many replays.

IIRC also POWER4 and its derivatives used a replay mechanism for handling memory operations which had missed in the caches. I do not know how much of a penalty it did pay for it but I guess it wasn't low as an instruction sent back for dispatch again had to be 'repackaged' in a new dispatch group.