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.
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.
JMP_FAR_I was unpacking its far pointer operand using sll instead of srl like
it should, and also putting the components in the wrong registers for use by
other microcode.
During iret access LDT/GDT at CPL0 rather than after transition to user mode
(if I'm reading the Intel IA-64 architecture spec correctly, the contents of
the descriptor table are read before the CPL is updated).
The code for Orion 2.0 makes use of printf() at several places where there as
an error in configuration of the model. These have been replaced with fatal().
By stalling and waiting the mandatory queue instead of recycling it, one can
ensure that no incoming messages are starved when the mandatory queue puts
signficant of pressure on the L1 cache controller (i.e. the ruby memtester).
--HG--
rename : src/mem/slicc/ast/WakeUpDependentsStatementAST.py => src/mem/slicc/ast/WakeUpAllDependentsStatementAST.py
The packet now identifies whether static or dynamic data has been allocated and
is used by Ruby to determine whehter to copy the data pointer into the ruby
request. Subsequently, Ruby can be told not to update phys memory when
receiving packets.
Move page table walker state to its own object type, and make the
walker instantiate state for each outstanding walk. By storing the
states in a queue, the walker is able to handle multiple outstanding
timing requests. Note that functional walks use separate state
elements.
In sendSplitData, keep a pointer to the senderState that may be updated after
the call to handle*Packet. This way, if the receiver updates the packet
senderState, it can still be accessed in sendSplitData.
Double packet delete problem is due to an interrupt device deleting a packet that the SimpleTimingPort also deletes. Since MessagePort descends from SimpleTimingPort, simply reimplement the failing code from SimpleTimingPort: recvTiming.
Separate data VCs and ctrl VCs in garnet, as ctrl VCs have 1 buffer per VC,
while data VCs have > 1 buffers per VC. This is for correct power estimations.
Maintain all information about an instruction's fault in the DynInst object rather
than any cpu-request object. Also, if there is a fault during the execution stage
then just save the fault inside the instruction and trap once the instruction
tries to graduate
Give fetch unit it's own parameterizable fetch buffer to read from. Very inefficient
(architecturally and in simulation) to continually fetch at the granularity of the
wordsize. As expected, the number of fetch memory requests drops dramatically
instead of having one cache-unit class be responsible for both data and code
accesses, separate code that is just for fetch in it's own derived class off the
original base class. This makes the code easier to manage as well as handle
future cases of special fetch handling
allow the user to specify how many instructions a pipeline stage can process
on any given cycle (stageWidth...i.e.bandwidth) by setting the parameter through
the python interface rather than compile the code after changing the *.cc file.
(we always had the parameter there, but still used the static 'ThePipeline::StageWidth'
instead)
-
Since StageWidth is now dynamically defined, change the interstage communication
structure to use a vector and get rid of array and array handling index (toNextStageIndex)
since we can just make calls to the list for the same information
use skidbuffer as only location for instructions between stages. before,
we had the insts queue from the prior stage and the skidbuffer for the
current stage, but that gets confusing and this consolidation helps
when handling squash cases
manage insertion and deletion like a queue but will need
access to internal elements for future changes
Currently, skidbuffer manages any instruction that was
in a stage but could not complete processing, however
we will want to manage all blocked instructions (from prev stage
and from cur. stage) in just one buffer.
Previous code was marking CPU activity on almost every cycle due to a bug in
tracking the status of pipeline stages. This disables the CPU from sleeping
on long latency stalls and increases simulation time
This makes sure that the address ranges requested for caches and uncached ports
don't conflict with each other, and that accesses which are always uncached
(message signaled interrupts for instance) don't waste time passing through
caches.
Moving the definition of NoFault into fault.hh doesn't bring any new
dependencies with it, and allows some files to include just fault.hh which has
less baggage. NoFault will still be available to everything that includes
faults.hh because it includes fault.hh.
M5 skips over any simulated time where it doesn't have any work to do. When
the simulation is active, the time skipped is short and the work done at any
point in time is relatively substantial. If the time between events is long
and/or the work to do at each event is small, it's possible for simulated time
to pass faster than real time. When running a benchmark that can be good
because it means the simulation will finish sooner in real time. When
interacting with the real world through, for instance, a serial terminal or
bridge to a real network, this can be a problem. Human or network response time
could be greatly exagerated from the perspective of the simulation and make
simulated events happen "too soon" from an external perspective.
This change adds the capability to force the simulation to run no faster than
real time. It does so by scheduling a periodic event that checks to see if
its simulated period is shorter than its real period. If it is, it stalls the
simulation until they're equal. This is called time syncing.
A future change could add pseudo instructions which turn time syncing on and
off from within the simulation. That would allow time syncing to be used for
the interactive parts of a session but then turned off when running a
benchmark using the m5 utility program inside a script. Time syncing would
probably not happen anyway while running a benchmark because there would be
plenty of work for M5 to do, but the event overhead could be avoided.
Any change of control flow now resets the itstate to 0 mask and 0 condition,
except where the control flow alteration write into the cpsr register. These
case, for example return from an iterrupt, require the predecoder to recover
the itstate.
As there is a window of opportunity between the return from an interrupt
changing the control flow at the head of the pipe and the commit of the update
to the CPSR, the predecoder needs to be able to grab the ITstate early. This
is now handled by setting the forcedItState inside a PCstate for the control
flow altering instruction.
That instruction will have the correct mask/cond, but will not have a valid
itstate until advancePC is called (note this happens to advance the execution).
When the new PCstate is copy constructed it gets the itstate cond/mask, and
upon advancing the PC the itstate becomes valid.
Subsequent advancing invalidates the state and zeroes the cond/mask. This is
handled in isolation for the ARM ISA and should have no impact on other ISAs.
Refer arch/arm/types.hh and arch/arm/predecoder.cc for the details.
Without this change 0 is always used for the youngest sequence number if
a squash occured and the ROB was empty (E.g. an instruction is marked
serializeAfter or a fetch stall prevents other instructions from issuing).
Using 0 there is a race to rename where an instruction that committed the
same cycle as the squashing instruction can have it's renamed state undone
by the squash using sequence number 0.
I'm not positive this is the correct fix, but it's working right now.
Either we need to do something like this, prevent the misc reg from being renamed at all,
or there something else going on. We need to find the root cause as to why
this is only a problem sometimes.
The squash inside the fetch unit should not attempt to remove them from the
branch predictor as non-control instructions are not pushed into the predictor.
When this condition occurs the cpu should restart the fetch stage to fetch from
the original execution path. Fault handling in the commit stage is cleaned up a
little bit so the control flow is simplier. Finally, if an instruction is being
used to carry a fault it isn't executed, so the fault propagates appropriately.
The purpose of this patch is to change the way CacheMemory interfaces with
coherence protocols. Currently, whenever a cache controller (defined in the
protocol under consideration) needs to carry out any operation on a cache
block, it looks up the tag hash map and figures out whether or not the block
exists in the cache. In case it does exist, the operation is carried out
(which requires another lookup). As observed through profiling of different
protocols, multiple such lookups take place for a given cache block. It was
noted that the tag lookup takes anything from 10% to 20% of the simulation
time. In order to reduce this time, this patch is being posted.
I have to acknowledge that the many of the thoughts that went in to this
patch belong to Brad.
Changes to CacheMemory, TBETable and AbstractCacheEntry classes:
1. The lookup function belonging to CacheMemory class now returns a pointer
to a cache block entry, instead of a reference. The pointer is NULL in case
the block being looked up is not present in the cache. Similar change has
been carried out in the lookup function of the TBETable class.
2. Function for setting and getting access permission of a cache block have
been moved from CacheMemory class to AbstractCacheEntry class.
3. The allocate function in CacheMemory class now returns pointer to the
allocated cache entry.
Changes to SLICC:
1. Each action now has implicit variables - cache_entry and tbe. cache_entry,
if != NULL, must point to the cache entry for the address on which the action
is being carried out. Similarly, tbe should also point to the transaction
buffer entry of the address on which the action is being carried out.
2. If a cache entry or a transaction buffer entry is passed on as an
argument to a function, it is presumed that a pointer is being passed on.
3. The cache entry and the tbe pointers received __implicitly__ by the
actions, are passed __explicitly__ to the trigger function.
4. While performing an action, set/unset_cache_entry, set/unset_tbe are to
be used for setting / unsetting cache entry and tbe pointers respectively.
5. is_valid() and is_invalid() has been made available for testing whether
a given pointer 'is not NULL' and 'is NULL' respectively.
6. Local variables are now available, but they are assumed to be pointers
always.
7. It is now possible for an object of the derieved class to make calls to
a function defined in the interface.
8. An OOD token has been introduced in SLICC. It is same as the NULL token
used in C/C++. If you are wondering, OOD stands for Out Of Domain.
9. static_cast can now taken an optional parameter that asks for casting the
given variable to a pointer of the given type.
10. Functions can be annotated with 'return_by_pointer=yes' to return a
pointer.
11. StateMachine has two new variables, EntryType and TBEType. EntryType is
set to the type which inherits from 'AbstractCacheEntry'. There can only be
one such type in the machine. TBEType is set to the type for which 'TBE' is
used as the name.
All the protocols have been modified to conform with the new interface.
This test exercises each of the functions in the reference counting pointer
implementation individually (except get()) and verifies they have some
minimially expected behavior. It also checks that reference counted objects
are freed when their usage count goes to 0 in some basic situations,
specifically a pointer being set to NULL and a pointer being deleted.
There's no reason for it to derive from SimLoopExitEvent.
This whole drain thing needs to be redone eventually,
but this is a stopgap to make later changes to
SimLoopExitEvent feasible.
Avoid direct references to mainEventQueue in pseudo-insts
by indirecting through associated CPU object.
Made exitSimLoop() more flexible to enable some of these.
There were several copies of similar functions that looked
like they all replicated reschedule(), so I replaced them
with direct calls. Keeping this separate from the previous
cset since there may be some subtle functional differences
if the code ever reschedules an event that is scheduled but
not squashed (though none were detected in the regressions).
Events need to be scheduled on the queue assigned
to the SimObject, not on the global queue (which
should be going away).
Also cleaned up a number of redundant expressions
that made the code unnecessarily verbose.
I like the brevity of Ali's recent change, but the ambiguity of
sometimes showing the source and sometimes the target is a little
confusing. This patch makes scons typically list all sources and
all targets for each action, with the common path prefix factored
out for brevity. It's a little more verbose now but also more
informative.
Somehow Ali talked me into adding colors too, which is a whole
'nother story.
This patch changes the manner in which data is copied from L1 to L2 cache in
the implementation of the Hammer's cache coherence protocol. Earlier, data was
copied directly from one cache entry to another. This has been broken in to
two parts. First, the data is copied from the source cache entry to a
transaction buffer entry. Then, data is copied from the transaction buffer
entry to the destination cache entry.
This has been done to maintain the invariant - at any given instant, multiple
caches under a controller are exclusive with respect to each other.
These files really aren't general enough to belong in src/base.
This patch doesn't reorder include lines, leaving them unsorted
in many cases, but Nate's magic script will fix that up shortly.
--HG--
rename : src/base/sched_list.hh => src/cpu/sched_list.hh
rename : src/base/timebuf.hh => src/cpu/timebuf.hh
Ran all the source files through 'perl -pi' with this script:
s|\s*(};?\s*)?/\*\s*(end\s*)?namespace\s*(\S+)\s*\*/(\s*})?|} // namespace $3|;
s|\s*};?\s*//\s*(end\s*)?namespace\s*(\S+)\s*|} // namespace $2\n|;
s|\s*};?\s*//\s*(\S+)\s*namespace\s*|} // namespace $1\n|;
Also did a little manual editing on some of the arch/*/isa_traits.hh files
and src/SConscript.
These operators were expecting a const T& instead of a const T*, and were not
being picked up and used by gcc in the right places as a result. Apparently no
one used these operators before. A unit test which exposed these problems,
verified the solution, and checks other basic functionality is on the way.
Two functions in src/mem/ruby/system/PerfectCacheMemory.hh, tryCacheAccess()
and cacheProbe(), end with calls to panic(). Both of these functions have
return type other than void. Any file that includes this header file fails
to compile because of the missing return statement. This patch adds dummy
values so as to avoid the compiler warnings.
This diff is for changing the way ASSERT is handled in Ruby. m5.fast
compiles out the assert statements by using the macro NDEBUG. Ruby uses the
macro RUBY_NO_ASSERT to do so. This macro has been removed and NDEBUG has
been put in its place.
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.
For SPARC ASIs are added to the ExtMachInst. If the ASI is changed simply
marking the instruction as Serializing isn't enough beacuse that only
stops rename. This provides a mechanism to squash all the instructions
and refetch them
ARM instructions updating cumulative flags (ARM FP exceptions and saturation
flags) are not serialized.
Added aliases for ARM FP exceptions and saturation flags in FPSCR. Removed
write accesses to the FP condition codes for most ARM VFP instructions: only
VCMP and VCMPE instructions update the FP condition codes. Removed a potential
cause of seg. faults in the O3 model for NEON memory macro-ops (ARM).
New parameter forms are:
IP address in the format "a.b.c.d" where a-d are from decimal 0 to 255.
IP address with netmask which is an IP followed by "/n" where n is a netmask
length in bits from decimal 0 to 32 or by "/e.f.g.h" where e-h are from
decimal 0 to 255 and which is all 1 bits followed by all 0 bits when
represented in binary. These can also be specified as an integral IP and
netmask passed in separately.
IP address with port which is an IP followed by ":p" where p is a port index
from decimal 0 to 65535. These can also be specified as an integral IP and
port value passed in separately.
This change makes O3 flatten floating point destination registers, and also
fixes misc register flattening so that it's correctly repositioned relative to
the resized regions for integer and floating point indices.
It also fixes some overly long lines.
In the case of a split transaction and a cache that is faster than a CPU we
could get two responses before next_tick expires. Add an event that is
scheduled in this case and return false rather than asserting.
The L1 cache may have been accessed to provide this data, which confuses
it, if it ends up being accesses twice in one cycle. Instead wait 1 tick
which will force the timing simple CPU to forward to its next clock cycle
when the translation completes.
Also prevent multiple outstanding table walks from occuring at once.
This change modifies the way prefetches work. They are now like normal loads
that don't writeback a register. Previously prefetches were supposed to call
prefetch() on the exection context, so they executed with execute() methods
instead of initiateAcc() completeAcc(). The prefetch() methods for all the CPUs
are blank, meaning that they get executed, but don't actually do anything.
On Alpha dead cache copy code was removed and prefetches are now normal ops.
They count as executed operations, but still don't do anything and IsMemRef is
not longer set on them.
On ARM IsDataPrefetch or IsInstructionPreftech is now set on all prefetch
instructions. The timing simple CPU doesn't try to do anything special for
prefetches now and they execute with the normal memory code path.
The build_dir parameter name has been deprecated and replaced with
variant_dir. This change switches us over to avoid warning spew in newer
versions of scons.
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.
Code in the CPUs that need a nop to carry a fault can't easily deal with a
microcoded nop. This instruction format provides for one that isn't.
--HG--
rename : src/arch/x86/isa/formats/syscall.isa => src/arch/x86/isa/formats/nop.isa
