Is it possible for the RESOURCE_STALLS.RS event to occur even when the RS is not completely full?

Question

The description of the RESOURCE_STALLS.RS hardware performance event for Intel Broadwell is the following:

This event counts stall cycles

Answer 1

Yes, it is possible for RESOURCE_STALLS to indicate a full RS before the RS is completely full.

As the RS becomes full, allocation of new uops into the RS becomes less ideal until at some point it may stall out entirely, even though some entries remain.

Furthermore, not all RS entries are available to all instructions. For example, on Haswell, I observe that only 30-32 of the 60 RS entries are available to loads: these entries may be special in they support uop replay, for example. On Skylake, the situation is different: the entire RS is not available to any type of instruction: rather, the "97 entry" RS is actually made up of a 64-entry RS for ALU ops, and a 33 entry RS for load ops. So the entire 97 entries of RS(es) will rarely be full, unless by some coincidence both fill up at exactly the same moment.

The RESOURCE_STALLS.RS event (umask 0x4) only triggers when a the "ALU" part of the RS is full (or full enough that an op can't allocate). For the load RS (which overlaps with the ALU RS in Haswell but not Skylake), the corresponding event has umask 0x40. You can use it with perf as 'cpu/event=0xa2,umask=0x40,name=resource_stalls_memrs_full/. Although the events are not documented for Skylake, they seem to work fine (although events with umasks 0x10 through 0x80 are very different than documented on Sandy Bridge.

Future Intel chips are likely to have even finer-grained reservation stations.