we are going to be getting away from creating new resource requests for every
instruction so no more need to keep track of a reqRemoveList and clean it up
every tick
first change in an optimization that will stop InOrder from allocating new memory for every instruction's
request to a resource. This gets expensive since every instruction needs to access ~10 requests before
graduation. Instead, the plan is to allocate just enough resource request objects to satisfy each resource's
bandwidth (e.g. the execution unit would need to allocate 3 resource request objects for a 1-issue pipeline
since on any given cycle it could have 2 read requests and 1 write request) and then let the instructions
contend and reuse those allocated requests. The end result is a smaller memory footprint for the InOrder model
and increased simulation performance
Currently the wakeup function for the PerfectSwitch contains three loops -
loop on number of virtual networks
loop on number of incoming links
loop till all messages for this (link, network) have been routed
With an 8 processor mesh network and Hammer protocol, about 11-12% of the
was observed to have been spent in this function, which is the highest
amongst all the functions. It was found that the innermost loop is executed
about 45 times per invocation of the wakeup function, when each invocation
of the wakeup function processes just about one message.
The patch tries to do away with the redundant executions of the innermost
loop. Counters have been added for each virtual network that record the
number of messages that need to be routed for that virtual network. The
inner loops are only executed when the number of messages for that particular
virtual network > 0. This does away with almost 80% of the executions of the
innermost loop. The function now consumes about 5-6% of the total execution
time.
In x86, 32 and 64 bit writes to registers in which registers appear to be 32 or
64 bits wide overwrite all bits of the destination register. This change
removes false dependencies in these cases where the previous value of a
register doesn't need to be read to write a new value. New versions of most
microops are created that have a "Big" suffix which simply overwrite their
destination, and the right version to use is selected during microop
allocation based on the selected data size.
This does not change the performance of the O3 CPU model significantly, I
assume because there are other false dependencies from the condition code bits
in the flags register.
These faults can panic/warn/warn_once, etc., instead of instructions doing
that themselves directly. That way, instructions can be speculatively
executed, and only if they're actually going to commit will their fault be
invoked and the panic, etc., happen.
When redirecting fetch to handle branches, the npc of the current pc state
needs to be left alone. This change makes the pc state record whether or not
the npc already reflects a real value by making it keep track of the current
instruction size, or if no size has been set.
The patch changes the order in which L1 dcache and icache are looked up when
a request comes in. Earlier, if a request came in for instruction fetch, the
dcache was looked up before the icache, to correctly handle self-modifying
code. But, in the common case, dcache is going to report a miss and the
subsequent icache lookup is going to report a hit. Given the invariant -
caches under the same controller keep track of disjoint sets of cache blocks,
we can move the icache lookup before the dcache lookup. In case of a hit in
the icache, using our invariant, we know that the dcache would have reported
a miss. In case of a miss in the icache, we know that icache would have
missed even if the dcache was looked up before looking up the icache.
Effectively, we are doing the same thing as before, though in the common case,
we expect reduction in the number of lookups. This was empirically confirmed
for MOESI hammer. The ratio lookups to access requests is now about 1.1 to 1.
there are still only a few inorder benchmark but for the lengthier benchmarks (twolf and vortext)
the latest changes to how instruction scheduling (how instructions figure out what they want to
do on each pipeline stage in the inorder model) were able to improve performance by a nice
amount... The latest results for the inorder model process about 100k insts/second
(note: 58% is over the last time run on 64-bit pool machines at UM)
resource skeds are divided into two parts: front end (all insts) and back end (inst. specific)
each of those are implemented as separate lists, so this iterator wraps around
the traditional list iterator so that an instruction can walk it's schedule but seamlessly
transfer from front end to back end when necessary
add a stage scheduler class to replace InstStage in pipeline_traits.cc
use that class to define a default front-end, resource schedule that all
instructions will follow. This will also replace the back end schedule in
pipeline_traits.cc. The reason for adding this is so that we can cache
instruction schedules in the future instead of calling the same function
over/over again as well as constantly dynamically alllocating memory on
every instruction to try to figure out it's schedule
When a table walk is initiated by the fetch stage, the CPU can
potentially move to the idle state and never wake up.
The fetch stage must call cpu->wakeCPU() when a translation completes
(in finishTranslation()).
Uncacheable requests were set as such only in atomic mode.
currState->delayed is checked in place of currState->timing for resetting
currState in atomic mode.
This change fixes an issue where a DTLB fault occurs and redirects fetch to
handle the fault and the ITLB requires a walk which delays translation. In this
case the status of the cpu isn't updated appropriately, and an additional
instruction fetch occurs. Eventually this hits an assert as multiple instruction
fetches are occuring in the system and when the second one returns the
processor is in the wrong state.
Some asserts below are removed because it was always true (typo) and the state
after the initiateAcc() the processor could be in any valid state when a
d-side fault occurs.
Some ISAs (like ARM) relies on hardware page table walkers. For those ISAs,
when a TLB miss occurs, initiateTranslation() can return with NoFault but with
the translation unfinished.
Instructions experiencing a delayed translation due to a hardware page table
walk are deferred until the translation completes and kept into the IQ. In
order to keep track of them, the IQ has been augmented with a queue of the
outstanding delayed memory instructions. When their translation completes,
instructions are re-executed (only their initiateAccess() was already
executed; their DTB translation is now skipped). The IEW stage has been
modified to support such a 2-pass execution.
Setup initial timesync event in initState or loadState so that curTick has
been updated to the new value, otherwise the event is scheduled in the past.
The TBE pointer in the MESI CMP implementation was not being set to NULL
when the TBE is deallocated. This resulted in segmentation fault on testing
the protocol when the ProtocolTrace was switched on.
