cachePorts currently constrains the number of store packets written to the
D-Cache each cycle), but loads currently affect this variable. This leads
to unexpected congestion (e.g., setting cachePorts to a realistic 1 will
in fact allow a store to WB only if no loads have accessed the D-Cache
this cycle). In the absence of arbitration, this patch decouples how many
loads can be done per cycle from how many stores can be done per cycle.
Signed-off-by: Jason Lowe-Power <jason@lowepower.com>
The read() function merely initiates a memory read operation; the
data doesn't arrive until the access completes and a response packet
is received from the memory system. Thus there's no need to provide
a data pointer; its existence is historical.
Getting this pointer out of this internal o3 interface sets the
stage for similar cleanup in the ExecContext interface. Also
found that we were pointlessly setting the contents at this pointer
on a store forward (the useful memcpy happens just a few lines
below the deleted one).
This patch moves away from using M5_ATTR_OVERRIDE and the m5::hashmap
(and similar) abstractions, as these are no longer needed with gcc 4.7
and clang 3.1 as minimum compiler versions.
The Request::UNCACHEABLE flag currently has two different
functions. The first, and obvious, function is to prevent the memory
system from caching data in the request. The second function is to
prevent reordering and speculation in CPU models.
This changeset gives the order/speculation requirement a separate flag
(Request::STRICT_ORDER). This flag prevents CPU models from doing the
following optimizations:
* Speculation: CPU models are not allowed to issue speculative
loads.
* Write combining: CPU models and caches are not allowed to merge
writes to the same cache line.
Note: The memory system may still reorder accesses unless the
UNCACHEABLE flag is set. It is therefore expected that the
STRICT_ORDER flag is combined with the UNCACHEABLE flag to prevent
this behavior.
Another churn to clean up undefined behaviour, mostly ARM, but some
parts also touching the generic part of the code base.
Most of the fixes are simply ensuring that proper intialisation. One
of the more subtle changes is the return type of the sign-extension,
which is changed to uint64_t. This is to avoid shifting negative
values (undefined behaviour) in the ISA code.
This patch takes quite a large step in transitioning from the ad-hoc
RefCountingPtr to the c++11 shared_ptr by adopting its use for all
Faults. There are no changes in behaviour, and the code modifications
are mostly just replacing "new" with "make_shared".
This patch fixes the load blocked/replay mechanism in the o3 cpu. Rather than
flushing the entire pipeline, this patch replays loads once the cache becomes
unblocked.
Additionally, deferred memory instructions (loads which had conflicting stores),
when replayed would not respect the number of functional units (only respected
issue width). This patch also corrects that.
Improvements over 20% have been observed on a microbenchmark designed to
exercise this behavior.
As highlighed on the mailing list gem5's writeback modeling can impact
performance. This patch removes the limitation on maximum outstanding issued
instructions, however the number that can writeback in a single cycle is still
respected in instToCommit().
Check for free entries in Load Queue and Store Queue separately to
avoid cases when load cannot be renamed due to full Store Queue and
vice versa.
This work was done while Binh was an intern at AMD Research.
In the O3 LSQ, data read/written is printed out in DPRINTFs. However,
the data field is treated as a character string with a null terminated.
However the data field is not encoded this way. This patch removes
that possibility by removing the data part of the print.
Previously, the O3 CPU could stop in the middle of a microcode
sequence. This patch makes sure that the pipeline stops when it has
committed a normal instruction or exited from a microcode
sequence. Additionally, it makes sure that the pipeline has no
instructions in flight when it is drained, which should make draining
more robust.
Draining is controlled in the commit stage, which checks if the next
PC after a committed instruction is in microcode. If this isn't the
case, it requests a squash of all instructions after that the
instruction that just committed and immediately signals a drain stall
to the fetch stage. The CPU then continues to execute until the
pipeline and all associated buffers are empty.
This patch addresses the comments and feedback on the preceding patch
that reworks the clocks and now more clearly shows where cycles
(relative cycle counts) are used to express time.
Instead of bumping the existing patch I chose to make this a separate
patch, merely to try and focus the discussion around a smaller set of
changes. The two patches will be pushed together though.
This changes done as part of this patch are mostly following directly
from the introduction of the wrapper class, and change enough code to
make things compile and run again. There are definitely more places
where int/uint/Tick is still used to represent cycles, and it will
take some time to chase them all down. Similarly, a lot of parameters
should be changed from Param.Tick and Param.Unsigned to
Param.Cycles.
In addition, the use of curTick is questionable as there should not be
an absolute cycle. Potential solutions can be built on top of this
patch. There is a similar situation in the o3 CPU where
lastRunningCycle is currently counting in Cycles, and is still an
absolute time. More discussion to be had in other words.
An additional change that would be appropriate in the future is to
perform a similar wrapping of Tick and probably also introduce a
Ticks class along with suitable operators for all these classes.
This patch introduces the notion of a clock update function that aims
to avoid costly divisions when turning the current tick into a
cycle. Each clocked object advances a private (hidden) cycle member
and a tick member and uses these to implement functions for getting
the tick of the next cycle, or the tick of a cycle some time in the
future.
In the different modules using the clocks, changes are made to avoid
counting in ticks only to later translate to cycles. There are a few
oddities in how the O3 and inorder CPU count idle cycles, as seen by a
few locations where a cycle is subtracted in the calculation. This is
done such that the regression does not change any stats, but should be
revisited in a future patch.
Another, much needed, change that is not done as part of this patch is
to introduce a new typedef uint64_t Cycle to be able to at least hint
at the unit of the variables counting Ticks vs Cycles. This will be
done as a follow-up patch.
As an additional follow up, the thread context still uses ticks for
the book keeping of last activate and last suspend and this should
probably also be changed into cycles as well.
This patch fixes some problems with the drain/switchout functionality
for the O3 cpu and for the ARM ISA and adds some useful debug print
statements.
This is an incremental fix as there are still a few bugs/mem leaks with the
switchout code. Particularly when switching from an O3CPU to a
TimingSimpleCPU. However, when switching from O3 to O3 cores with the ARM ISA
I haven't encountered any more assertion failures; now the kernel will
typically panic inside of simulation.
While FastAlloc provides a small performance increase (~1.5%) over regular malloc it isn't thread safe.
After removing FastAlloc and using tcmalloc I've seen a performance increase of 12% over libc malloc
when running twolf for ARM.
This patch moves send/recvTiming and send/recvTimingSnoop from the
Port base class to the MasterPort and SlavePort, and also splits them
into separate member functions for requests and responses:
send/recvTimingReq, send/recvTimingResp, and send/recvTimingSnoopReq,
send/recvTimingSnoopResp. A master port sends requests and receives
responses, and also receives snoop requests and sends snoop
responses. A slave port has the reciprocal behaviour as it receives
requests and sends responses, and sends snoop requests and receives
snoop responses.
For all MemObjects that have only master ports or slave ports (but not
both), e.g. a CPU, or a PIO device, this patch merely adds more
clarity to what kind of access is taking place. For example, a CPU
port used to call sendTiming, and will now call
sendTimingReq. Similarly, a response previously came back through
recvTiming, which is now recvTimingResp. For the modules that have
both master and slave ports, e.g. the bus, the behaviour was
previously relying on branches based on pkt->isRequest(), and this is
now replaced with a direct call to the apprioriate member function
depending on the type of access. Please note that send/recvRetry is
still shared by all the timing accessors and remains in the Port base
class for now (to maintain the current bus functionality and avoid
changing the statistics of all regressions).
The packet queue is split into a MasterPort and SlavePort version to
facilitate the use of the new timing accessors. All uses of the
PacketQueue are updated accordingly.
With this patch, the type of packet (request or response) is now well
defined for each type of access, and asserts on pkt->isRequest() and
pkt->isResponse() are now moved to the appropriate send member
functions. It is also worth noting that sendTimingSnoopReq no longer
returns a boolean, as the semantics do not alow snoop requests to be
rejected or stalled. All these assumptions are now excplicitly part of
the port interface itself.
This patch simplifies the packet by removing the broadcast flag and
instead more firmly relying on (and enforcing) the semantics of
transactions in the classic memory system, i.e. request packets are
routed from a master to a slave based on the address, and when they
are created they have neither a valid source, nor destination. On
their way to the slave, the request packet is updated with a source
field for all modules that multiplex packets from multiple master
(e.g. a bus). When a request packet is turned into a response packet
(at the final slave), it moves the potentially populated source field
to the destination field, and the response packet is routed through
any multiplexing components back to the master based on the
destination field.
Modules that connect multiplexing components, such as caches and
bridges store any existing source and destination field in the sender
state as a stack (just as before).
The packet constructor is simplified in that there is no longer a need
to pass the Packet::Broadcast as the destination (this was always the
case for the classic memory system). In the case of Ruby, rather than
using the parameter to the constructor we now rely on setDest, as
there is already another three-argument constructor in the packet
class.
In many places where the packet information was printed as part of
DPRINTFs, request packets would be printed with a numeric "dest" that
would always be -1 (Broadcast) and that field is now removed from the
printing.
This patch introduces the notion of a master and slave port in the C++
code, thus bringing the previous classification from the Python
classes into the corresponding simulation objects and memory objects.
The patch enables us to classify behaviours into the two bins and add
assumptions and enfore compliance, also simplifying the two
interfaces. As a starting point, isSnooping is confined to a master
port, and getAddrRanges to slave ports. More of these specilisations
are to come in later patches.
The getPort function is not getMasterPort and getSlavePort, and
returns a port reference rather than a pointer as NULL would never be
a valid return value. The default implementation of these two
functions is placed in MemObject, and calls fatal.
The one drawback with this specific patch is that it requires some
code duplication, e.g. QueuedPort becomes QueuedMasterPort and
QueuedSlavePort, and BusPort becomes BusMasterPort and BusSlavePort
(avoiding multiple inheritance). With the later introduction of the
port interfaces, moving the functionality outside the port itself, a
lot of the duplicated code will disappear again.
Because there are no longer architecture independent but specialized functions
in arch/XXX/faults.hh, code that isn't using the faults from a particular ISA
no longer needs to be able to include them through the switching header file
arch/faults.hh. By removing that header file (arch/faults.hh), the potential
interface between ISA code and non ISA code is narrowed.
This patch adds the necessary flags to the SConstruct and SConscript
files for compiling using clang 2.9 and later (on Ubuntu et al and OSX
XCode 4.2), and also cleans up a bunch of compiler warnings found by
clang. Most of the warnings are related to hidden virtual functions,
comparisons with unsigneds >= 0, and if-statements with empty
bodies. A number of mismatches between struct and class are also
fixed. clang 2.8 is not working as it has problems with class names
that occur in multiple namespaces (e.g. Statistics in
kernel_stats.hh).
clang has a bug (http://llvm.org/bugs/show_bug.cgi?id=7247) which
causes confusion between the container std::set and the function
Packet::set, and this is currently addressed by not including the
entire namespace std, but rather selecting e.g. "using std::vector" in
the appropriate places.
This patch makes O3's LSQ maintain total order between stores. Essentially
only the store at the head of the store buffer is allowed to be in flight.
Only after that store completes, the next store is issued to the memory
system. By default, the x86 architecture will have TSO.
Only create a memory ordering violation when the value could have changed
between two subsequent loads, instead of just when loads go out-of-order
to the same address. While not very common in the case of Alpha, with
an architecture with a hardware table walker this can happen reasonably
frequently beacuse a translation will miss and start a table walk and
before the CPU re-schedules the faulting instruction another one will
pass it to the same address (or cache block depending on the dendency
checking).
This patch has been tested with a couple of self-checking hand crafted
programs to stress ordering between two cores.
The performance improvement on SPEC benchmarks can be substantial (2-10%).
Calculation of offset to copy from storeQueue[idx].data structure for load to
store forwarding fixed to be difference in bytes between store and load virtual
addresses. Previous method would induce bug where a load would index into
buffer at the wrong location.
If a split load fails on a blocked cache wbOutstanding can be decremented
twice if the first part of the split load succeeds and the second part fails.
Condition the decrementing on not having completed the first part of the load.
At the same time, rename the trace flags to debug flags since they
have broader usage than simply tracing. This means that
--trace-flags is now --debug-flags and --trace-help is now --debug-help
The comment in the code suggests that the checking granularity should be 16
bytes, however in reality the shift by 8 is 256 bytes which seems much
larger than required.
The store queue doesn't need to be ISA specific and architectures can
frequently store more than an int registers worth of data. A 128 bits seems
more common, but even 256 bits may be appropriate. Pretty much anything less
than a cache line size is buildable.
This change is a low level and pervasive reorganization of how PCs are managed
in M5. Back when Alpha was the only ISA, there were only 2 PCs to worry about,
the PC and the NPC, and the lsb of the PC signaled whether or not you were in
PAL mode. As other ISAs were added, we had to add an NNPC, micro PC and next
micropc, x86 and ARM introduced variable length instruction sets, and ARM
started to keep track of mode bits in the PC. Each CPU model handled PCs in
its own custom way that needed to be updated individually to handle the new
dimensions of variability, or, in the case of ARMs mode-bit-in-the-pc hack,
the complexity could be hidden in the ISA at the ISA implementation's expense.
Areas like the branch predictor hadn't been updated to handle branch delay
slots or micropcs, and it turns out that had introduced a significant (10s of
percent) performance bug in SPARC and to a lesser extend MIPS. Rather than
perpetuate the problem by reworking O3 again to handle the PC features needed
by x86, this change was introduced to rework PC handling in a more modular,
transparent, and hopefully efficient way.
PC type:
Rather than having the superset of all possible elements of PC state declared
in each of the CPU models, each ISA defines its own PCState type which has
exactly the elements it needs. A cross product of canned PCState classes are
defined in the new "generic" ISA directory for ISAs with/without delay slots
and microcode. These are either typedef-ed or subclassed by each ISA. To read
or write this structure through a *Context, you use the new pcState() accessor
which reads or writes depending on whether it has an argument. If you just
want the address of the current or next instruction or the current micro PC,
you can get those through read-only accessors on either the PCState type or
the *Contexts. These are instAddr(), nextInstAddr(), and microPC(). Note the
move away from readPC. That name is ambiguous since it's not clear whether or
not it should be the actual address to fetch from, or if it should have extra
bits in it like the PAL mode bit. Each class is free to define its own
functions to get at whatever values it needs however it needs to to be used in
ISA specific code. Eventually Alpha's PAL mode bit could be moved out of the
PC and into a separate field like ARM.
These types can be reset to a particular pc (where npc = pc +
sizeof(MachInst), nnpc = npc + sizeof(MachInst), upc = 0, nupc = 1 as
appropriate), printed, serialized, and compared. There is a branching()
function which encapsulates code in the CPU models that checked if an
instruction branched or not. Exactly what that means in the context of branch
delay slots which can skip an instruction when not taken is ambiguous, and
ideally this function and its uses can be eliminated. PCStates also generally
know how to advance themselves in various ways depending on if they point at
an instruction, a microop, or the last microop of a macroop. More on that
later.
Ideally, accessing all the PCs at once when setting them will improve
performance of M5 even though more data needs to be moved around. This is
because often all the PCs need to be manipulated together, and by getting them
all at once you avoid multiple function calls. Also, the PCs of a particular
thread will have spatial locality in the cache. Previously they were grouped
by element in arrays which spread out accesses.
Advancing the PC:
The PCs were previously managed entirely by the CPU which had to know about PC
semantics, try to figure out which dimension to increment the PC in, what to
set NPC/NNPC, etc. These decisions are best left to the ISA in conjunction
with the PC type itself. Because most of the information about how to
increment the PC (mainly what type of instruction it refers to) is contained
in the instruction object, a new advancePC virtual function was added to the
StaticInst class. Subclasses provide an implementation that moves around the
right element of the PC with a minimal amount of decision making. In ISAs like
Alpha, the instructions always simply assign NPC to PC without having to worry
about micropcs, nnpcs, etc. The added cost of a virtual function call should
be outweighed by not having to figure out as much about what to do with the
PCs and mucking around with the extra elements.
One drawback of making the StaticInsts advance the PC is that you have to
actually have one to advance the PC. This would, superficially, seem to
require decoding an instruction before fetch could advance. This is, as far as
I can tell, realistic. fetch would advance through memory addresses, not PCs,
perhaps predicting new memory addresses using existing ones. More
sophisticated decisions about control flow would be made later on, after the
instruction was decoded, and handed back to fetch. If branching needs to
happen, some amount of decoding needs to happen to see that it's a branch,
what the target is, etc. This could get a little more complicated if that gets
done by the predecoder, but I'm choosing to ignore that for now.
Variable length instructions:
To handle variable length instructions in x86 and ARM, the predecoder now
takes in the current PC by reference to the getExtMachInst function. It can
modify the PC however it needs to (by setting NPC to be the PC + instruction
length, for instance). This could be improved since the CPU doesn't know if
the PC was modified and always has to write it back.
ISA parser:
To support the new API, all PC related operand types were removed from the
parser and replaced with a PCState type. There are two warts on this
implementation. First, as with all the other operand types, the PCState still
has to have a valid operand type even though it doesn't use it. Second, using
syntax like PCS.npc(target) doesn't work for two reasons, this looks like the
syntax for operand type overriding, and the parser can't figure out if you're
reading or writing. Instructions that use the PCS operand (which I've
consistently called it) need to first read it into a local variable,
manipulate it, and then write it back out.
Return address stack:
The return address stack needed a little extra help because, in the presence
of branch delay slots, it has to merge together elements of the return PC and
the call PC. To handle that, a buildRetPC utility function was added. There
are basically only two versions in all the ISAs, but it didn't seem short
enough to put into the generic ISA directory. Also, the branch predictor code
in O3 and InOrder were adjusted so that they always store the PC of the actual
call instruction in the RAS, not the next PC. If the call instruction is a
microop, the next PC refers to the next microop in the same macroop which is
probably not desirable. The buildRetPC function advances the PC intelligently
to the next macroop (in an ISA specific way) so that that case works.
Change in stats:
There were no change in stats except in MIPS and SPARC in the O3 model. MIPS
runs in about 9% fewer ticks. SPARC runs with 30%-50% fewer ticks, which could
likely be improved further by setting call/return instruction flags and taking
advantage of the RAS.
TODO:
Add != operators to the PCState classes, defined trivially to be !(a==b).
Smooth out places where PCs are split apart, passed around, and put back
together later. I think this might happen in SPARC's fault code. Add ISA
specific constructors that allow setting PC elements without calling a bunch
of accessors. Try to eliminate the need for the branching() function. Factor
out Alpha's PAL mode pc bit into a separate flag field, and eliminate places
where it's blindly masked out or tested in the PC.
When each load or store is sent to the LSQ, we check whether it will cross a
cache line boundary and, if so, split it in two. This creates two TLB
translations and two memory requests. Care has to be taken if the first
packet of a split load is sent but the second blocks the cache. Similarly,
for a store, if the first packet cannot be sent, we must store the second
one somewhere to retry later.
This modifies the LSQSenderState class to record both packets in a split
load or store.
Finally, a new const variable, HasUnalignedMemAcc, is added to each ISA
to indicate whether unaligned memory accesses are allowed. This is used
throughout the changed code so that compiler can optimise away code dealing
with split requests for ISAs that don't need them.
