The x87 FPU supports three floating point formats: 32-bit, 64-bit, and
80-bit floats. The current gem5 implementation supports 32-bit and
64-bit floats, but only works correctly for 64-bit floats. This
changeset fixes the 32-bit float handling by correctly loading and
rounding (using truncation) 32-bit floats instead of simply truncating
the bit pattern.
80-bit floats are loaded by first loading the 80-bits of the float to
two temporary integer registers. A micro-op (cvtint_fp80) then
converts the contents of the two integer registers to the internal FP
representation (double). Similarly, when storing an 80-bit float,
there are two conversion routines (ctvfp80h_int and cvtfp80l_int) that
convert an internal FP register to 80-bit and stores the upper 64-bits
or lower 32-bits to an integer register, which is the written to
memory using normal integer stores.
X87 store instructions typically loads and pops the top value of the
stack and stores it in memory. The current implementation pops the
stack at the same time as the floating point value is loaded to a
temporary register. This will corrupt the state of the x87 stack if
the store fails. This changeset introduces a pop87 micro-instruction
that pops the stack and uses this instruction in the affected
macro-instructions to pop the stack after storing the value to memory.
The x87 FPU on x86 supports extended floating point. We currently
handle all floating point on x86 as double and don't support 80-bit
loads/stores. This changeset add a utility function to load and
convert 80-bit floats to doubles (loadFloat80) and another function to
store doubles as 80-bit floats (storeFloat80). Both functions use
libfputils to do the conversion in software. The functions are
currently not used, but are required to handle floating point in KVM
and to properly support all x87 loads/stores.
This changeset adds the convX87XTagsToTags() and convX87TagsToXTags()
which convert between the tag formats in the FTW register and the
format used in the xsave area. The conversion from to the x87 FTW
representation is currently loses some information since it does not
reconstruct the valid/zero/special flags which are not included in the
xsave representation.
In order to support hardware virtualization, we need to be able to
check if there are any interrupts pending irregardless of the
rflags.intf value. This changeset adds the checkInterruptsRaw() method
to the x86 interrupt control. It returns true if there are pending
interrupts that can be delivered as soon as the CPU is ready for
interrupt delivery.
This patch adds checkpointing support to x86 tlb. It upgrades the
cpt_upgrader.py script so that previously created checkpoints can
be updated. It moves the checkpoint version to 6.
Instead of relying on derived classes explicitly assigning
to the BasicPioDevice pioSize field, require them to pass
a size value in to the constructor.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
PciDev and IntDev stuck out as the only device classes that
ended in 'Dev' rather than 'Device'. This patch takes care
of that inconsistency.
Note that you may need to delete pre-existing files matching
build/*/python/m5/internal/param_* as scons does not pick up
indirect dependencies on imported python modules when generating
params, and the PciDev -> PciDevice rename takes place in a
file (dev/Device.py) that gets imported quite a bit.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
A couple of devices that have single fixed memory mapped regions
were not derived from BasicPioDevice, when that's exactly
the functionality that BasicPioDevice provides. This patch
gets rid of a little bit of redundant code by making those
devices actually do so.
Also fixed the weird case of X86ISA::Interrupts, where
the class already did derive from BasicPioDevice but
didn't actually use all the features it could have.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
The current implementation of the x87 never updates the x87 tag
word. This is currently not a big issue since the simulated x87 never
checks for stack overflows, however this becomes an issue when
switching between a virtualized CPU and a simulated CPU. This
changeset adds support, which is enabled by default, for updating the
tag register to every floating point microop that updates the stack
top using the spm mechanism.
The new tag words is generated by the helper function
X86ISA::genX87Tags(). This function is currently limited to flagging a
stack position as valid or invalid and does not try to distinguish
between the valid, zero, and special states.
This changeset actually fixes two issues:
* The lfpimm instruction didn't work correctly when applied to a
floating point constant (it did work for integers containing the
bit string representation of a constant) since it used
reinterpret_cast to convert a double to a uint64_t. This caused a
compilation error, at least, in gcc 4.6.3.
* The instructions loading floating point constants in the x87
processor didn't work correctly since they just stored a truncated
integer instead of a double in the floating point register. This
changeset fixes the old microcode by using lfpimm instruction
instead of the limm instructions.
The current implementation of fprem simply does an fmod and doesn't
simulate any of the iterative behavior in a real fprem. This isn't
normally a problem, however, it can lead to problems when switching
between CPU models. If switching from a real CPU in the middle of an
fprem loop to a simulated CPU, the output of the fprem loop becomes
correupted. This changeset changes the fprem implementation to work
like the one on real hardware.
The rflags register is spread across several different registers. Most
of the flags are stored in MISCREG_RFLAGS, but some are stored in
microcode registers. When accessing RFLAGS, we need to reconstruct it
from these registers. This changeset adds two functions,
X86ISA::getRFlags() and X86ISA::setRFlags(), that take care of this
magic.
This changeset fixes two problems in the FABS and FCHS
implementation. First, the ISA parser expects the assignment in
flag_code to be a pure assignment and not an and-assignment, which
leads to the isa_parser omitting the misc reg update. Second, the FCHS
and FABS macro-ops don't set the SetStatus flag, which means that the
default micro-op version, which doesn't update FSW, is executed.
The TSC value stored in MISCREG_TSC is actually just an offset from
the current CPU cycle to the actual TSC value. Writes with
side-effects to the TSC subtract the current cycle count before
storing the new value, while reads add the current cycle count. When
switching CPUs, the current value is copied without side-effects. This
works as long as the source and the destination CPUs have the same
clock frequencies. The TSC will jump, sometimes backwards, if they
have different clock frequencies. Most OSes assume the TSC to be
monotonic and break when this happens.
This changeset makes sure that the TSC is copied with side-effects to
ensure that the offset is updated to match the new CPU.
in the TLB
Some architectures (currently only x86) require some fixing-up of
physical addresses after a normal address translation. This is usually
to remap devices such as the APIC, but could be used for other memory
mapped devices as well. When running the CPU in a using hardware
virtualization, we still need to do these address fix-ups before
inserting the request into the memory system. This patch moves this
patch allows that code to be used by such CPUs without doing full
address translations.
This is the x86 version of the ARM changeset baa17ba80e06. In case an
instruction has been squashed by the o3 cpu, this patch allows page
table walker to avoid carrying out a pending translation that the
instruction requested for.
Currently call and return instructions are marked as IsCall and IsReturn. Thus, the
branch predictor does not use RAS for these instructions. Similarly, the number of
function calls that took place is recorded as 0. This patch marks these instructions
as they should be.
Currently all the integer microops are marked as IntAluOp and the floating
point microops are marked as FloatAddOp. This patch adds support for marking
different microops differently. Now IntMultOp, IntDivOp, FloatDivOp,
FloatMultOp, FloatCvtOp, FloatSqrtOp classes will be used as well. This will
help in providing different latencies for different op class.
The vsyscall address for gettimeofday is 0xffffffffff600000ul. The offset
therefore should be 0x0 instead of 0x410. This can be cross checked with
the file sysdeps/unix/sysv/linux/x86_64/gettimeofday.c in source of glibc.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
The 'lret' instruction reloads instruction pointer and code segment from the
stack and then pops them. But the popping part is missing from the current
implementation. This caused incorrect behavior in some code related to the
Fiasco OS. Microops are being added to rectify the behavior of the instruction.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
It is possible that operating system wants to shutdown the
lapic timer by writing timer's initial count to 0. This patch
adds a check that the timer event is only scheduled if the
count is 0.
The patch also converts few of the panics related to the keyboard
to warnings since we are any way not interested in simulating the
keyboard.
This patch fixes an issue related to the table walker recycling
packets that still have a bus delay that is not accounted for. For
now, we simply ignore the values and reset them to zero.
This patch fixes the warnings that clang3.2svn emit due to the "-Wall"
flag. There is one case of an uninitialised value in the ARM neon ISA
description, and then a whole range of unused private fields that are
pruned.
A derived function with a different signature than a base class
function will result in the base class function of the same name being
hidden. The parameter list and return type for the member function in
the derived class must match those of the member function in the base
class, otherwise the function in the derived class will hide the
function in the base class and no polymorphic behaviour will occur.
This patch addresses these warnings by ensuring a unique function name
to avoid (unintentionally) hiding any functions.
This patch moves the 16x APIC clock divider to the Python code to
avoid the post-instantiation modifications to the clock. The x86 APIC
was the only object setting the clock after creation time and this
required some custom functionality and configuration. With this patch,
the clock multiplier is moved to the Python code and the objects are
instantiated with the appropriate clock.
This patch adds a predecessor field to the SenderState base class to
make the process of linking them up more uniform, and enable a
traversal of the stack without knowing the specific type of the
subclasses.
There are a number of simplifications done as part of changing the
SenderState, particularly in the RubyTest.
Virtualized CPUs and the fastmem mode of the atomic CPU require direct
access to physical memory. We currently require caches to be disabled
when using them to prevent chaos. This is not ideal when switching
between hardware virutalized CPUs and other CPU models as it would
require a configuration change on each switch. This changeset
introduces a new version of the atomic memory mode,
'atomic_noncaching', where memory accesses are inserted into the
memory system as atomic accesses, but bypass caches.
To make memory mode tests cleaner, the following methods are added to
the System class:
* isAtomicMode() -- True if the memory mode is 'atomic' or 'direct'.
* isTimingMode() -- True if the memory mode is 'timing'.
* bypassCaches() -- True if caches should be bypassed.
The old getMemoryMode() and setMemoryMode() methods should never be
used from the C++ world anymore.
The changes made by the changeset 270c9a75e91f do not work well with switching
of cpus. The problem is that decoder for the old thread context holds state
that is not taken over by the new decoder.
This patch adds a takeOverFrom() function to Decoder class in each ISA. Except
for x86, functions in other ISAs are blank. For x86, the function copies state
from the old decoder to the new decoder.
Note that clflush is only being enabled. It is not implemented
in actual. A warning is printed if the cpu encounters a clflush
instruction. We need to enable this instruction in cpuid since
JRE 1.7 tests for it.
The changes made by the changeset 9376 were not quite correct. The patch made
changes to the code which resulted in decoder not getting initialized correctly
when the state was restored from a checkpoint.
This patch adds a startup function to each ISA object. For x86, this function
sets the required state in the decoder. For other ISAs, the function is empty
right now.
Used as a command in full-system scripts helps the user ensure the benchmarks have finished successfully.
For example, one can use:
/path/to/benchmark args || /sbin/m5 fail 1
and thus ensure gem5 will exit with an error if the benchmark fails.
After making the ISA an independent SimObject, it is serialized
automatically by the Python world. Previously, this just resulted in
an empty ISA section. This patch moves the contents of the ISA to that
section and removes the explicit ISA serialization from the thread
contexts, which makes it behave like a normal SimObject during
serialization.
Note: This patch breaks checkpoint backwards compatibility! Use the
cpt_upgrader.py utility to upgrade old checkpoints to the new format.
This patch adds support for the memInvalidate() drain method. TLB
flushing is requested by calling the virtual flushAll() method on the
TLB.
Note: This patch renames invalidateAll() to flushAll() on x86 and
SPARC to make the interface consistent across all supported
architectures.
The ISA class on stores the contents of ID registers on many
architectures. In order to make reset values of such registers
configurable, we make the class inherit from SimObject, which allows
us to use the normal generated parameter headers.
This patch introduces a Python helper method, BaseCPU.createThreads(),
which creates a set of ISAs for each of the threads in an SMT
system. Although it is currently only needed when creating
multi-threaded CPUs, it should always be called before instantiating
the system as this is an obvious place to configure ID registers
identifying a thread/CPU.
This patch unlocks the cpu-local monitor when the CPU sees a snoop to a locked
address. Previously we relied on the cache to handle the locking for us, however
some users on the gem5 mailing list reported a case where the cpu speculatively
executes a ll operation after a pending sc operation in the pipeline and that
makes the cache monitor valid. This should handle that case by invaliding the
local monitor.
This interface is no longer used, and getting rid of it simplifies the
decoders and code that sets up the decoders. The thread context had been used
to read architectural state which was used to contextualize the instruction
memory as it came in. That was changed so that the state is now sent to the
decoders to keep locally if/when it changes. That's significantly more
efficient.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
The predecoder in x86 does a lot of work, most of which can be skipped if the
decoder cache is put in front of it.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
This patch implements the fnstsw instruction. The code was originally written
by Vince Weaver. Gabe had made some comments about the code, but those were
never addressed. This patch addresses those comments.
This patch implements the fsincos instruction. The code was originally written
by Vince Weaver. Gabe had made some comments about the code, but those were
never addressed. This patch addresses those comments.
When casting objects in the generated SWIG interfaces, SWIG uses
classical C-style casts ( (Foo *)bar; ). In some cases, this can
degenerate into the equivalent of a reinterpret_cast (mainly if only a
forward declaration of the type is available). This usually works for
most compilers, but it is known to break if multiple inheritance is
used anywhere in the object hierarchy.
This patch introduces the cxx_header attribute to Python SimObject
definitions, which should be used to specify a header to include in
the SWIG interface. The header should include the declaration of the
wrapped object. We currently don't enforce header the use of the
header attribute, but a warning will be generated for objects that do
not use it.
This patch takes the Linux thread info support scattered across
different ISA implementations (currently in ARM, ALPHA, and MIPS), and
unifies them into a single file.
Adds a few more helper functions to read out TGID, mm, etc.
ISA-specific information (e.g., ALPHA PCBB register) is now moved to
the corresponding isa_traits.hh files.
This patch adds an additional level of ports in the inheritance
hierarchy, separating out the protocol-specific and protocl-agnostic
parts. All the functionality related to the binding of ports is now
confined to use BaseMaster/BaseSlavePorts, and all the
protocol-specific parts stay in the Master/SlavePort. In the future it
will be possible to add other protocol-specific implementations.
The functions used in the binding of ports, i.e. getMaster/SlavePort
now use the base classes, and the index parameter is updated to use
the PortID typedef with the symbolic InvalidPortID as the default.
This patch changes how the serialization of the system works. The base
class had a non-virtual serialize and unserialize, that was hidden by
a function with the same name for a number of subclasses (most likely
not intentional as the base class should have been virtual). A few of
the derived systems had no specialization at all (e.g. Power and x86
that simply called the System::serialize), but MIPS and Alpha adds
additional symbol table entries to the checkpoint.
Instead of overriding the virtual function, the additional entries are
now printed through a virtual function (un)serializeSymtab. The reason
for not calling System::serialize from the two related systems is that
a follow up patch will require the system to also serialize the
PhysicalMemory, and if this is done in the base class if ends up being
between the general parts and the specialized symbol table.
With this patch, the checkpoint is not modified, as the order of the
segments is unchanged.
This patch takes the final plunge and transitions from the templated
Range class to the more specific AddrRange. In doing so it changes the
obvious Range<Addr> to AddrRange, and also bumps the range_map to be
AddrRangeMap.
In addition to the obvious changes, including the removal of redundant
includes, this patch also does some house keeping in preparing for the
introduction of address interleaving support in the ranges. The Range
class is also stripped of all the functionality that is never used.
--HG--
rename : src/base/range.hh => src/base/addr_range.hh
rename : src/base/range_map.hh => src/base/addr_range_map.hh
The patch introduces two predicates for condition code registers -- one
tests if a register needs to be read, the other tests whether a register
needs to be written to. These predicates are evaluated twice -- during
construction of the microop and during its execution. Register reads
and writes are elided depending on how the predicates evaluate.
The D flag bit is part of the cc flag bit register currently. But since it
is not being used any where in the implementation, it creates an unnecessary
dependency. Hence, it is being moved to a separate register.
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 removes the NACK frrom the packet as there is no longer any
module in the system that issues them (the bridge was the only one and
the previous patch removes that).
The handling of NACKs was mostly avoided throughout the code base, by
using e.g. panic or assert false, but in a few locations the NACKs
were actually dealt with (although NACKs never occured in any of the
regressions). Most notably, the DMA port will now never receive a NACK
and the backoff time is thus never changed. As a consequence, the
entire backoff mechanism (similar to a PCI bus) is now removed and the
DMA port entirely relies on the bus performing the arbitration and
issuing a retry when appropriate. This is more in line with e.g. PCIe.
Surprisingly, this patch has no impact on any of the regressions. As
mentioned in the patch that removes the NACK from the bridge, a
follow-up patch should change the request and response buffer size for
at least one regression to also verify that the system behaves as
expected when the bridge fills up.
This patch removes the overloading of the parameter, which seems both
redundant, and possibly incorrect.
The PciConfigAll now also uses a Param.Latency rather than a
Param.Tick. For backwards compatibility it still sets the pio_latency
to 1 tick. All the comments have also been updated to not state that
it is in simticks when it is not necessarily the case.
This patch moves the clock of the CPU, bus, and numerous devices to
the new class ClockedObject, that sits in between the SimObject and
MemObject in the class hierarchy. Although there are currently a fair
amount of MemObjects that do not make use of the clock, they
potentially should do so, e.g. the caches should at some point have
the same clock as the CPU, potentially with a 1:n ratio. This patch
does not introduce any new clock objects or object hierarchies
(clusters, clock domains etc), but is still a step in the direction of
having a more structured approach clock domains.
The most contentious part of this patch is the serialisation of clocks
that some of the modules (but not all) did previously. This
serialisation should not be needed as the clock is set through the
parameters even when restoring from the checkpoint. In other words,
the state is "stored" in the Python code that creates the modules.
The nextCycle methods are also simplified and the clock phase
parameter of the CPU is removed (this could be part of a clock object
once they are introduced).
New tool chains seem to be looking for kernel versions newer than what
this this was previously set to. Also take this opportunity to change
the hostname we report in uname to sim.gem5.org.
Enable different whitelists for different OS/arch combinations,
since some use the generic Linux definitions only, and others
use definitions inherited from earlier Unix flavors on those
architectures.
Also update x86 function pointers so ioctl is no longer
unimplemented on that platform.
This patch is a revised version of Vince Weaver's earlier patch.
This patch makes getAddrRanges const throughout the code base. There
is no reason why it should not be, and making it const prevents adding
any unintentional side-effects.
Due to recent changes to X86 TLB, gem5 stopped compiling on
gcc version 4.4.3. This patch provides the fix for that problem. The patch
is tested on gcc 4.4.3. The change is not required for more recent
versions of gcc (like on 4.6.3).
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.
The CPUID instruction was implemented so that it would only write its results
if the instruction was successful. This works fine on the simple CPU where
unwritten registers retain their old values, but on a CPU like O3 with
renaming this is broken. The instruction needs to write the old values back
into the registers explicitly if they aren't being changed.
There are some bits of some fields of the ExtMachInst which are not actually
used for anything but are included in the hash of an ExtMachInst for
simplicity and efficiency. This change makes sure the decoder's internal
working ExtMachInst is completely initialized, even these unused bits, so that
there isn't any nondeterministic behavior, no valgrind messages about
uninitialized variables, and no potential false misses/redundant entries in
the decode cache.
The GDT can be accessed by user level software running in compatibility mode
by moving segment selectors into segment registers. The GDT needs to be set up
at an address accessible in this mode.
A small change was added a while ago to keep addresses from overflowing 32
bits when larger addresses shouldn't be accessible to software. That change
truncated when not in long mode, but really it should have truncated when not
in 64 bit mode. The difference is whether compatibility mode is included, a
mode that's supposed to act like a legacy 32 bit mode.
This will allow it to be specialized by the ISAs. The existing caching scheme
is provided by the BasicDecodeCache in the GenericISA namespace and is built
from the generalized components.
--HG--
rename : src/cpu/decode_cache.cc => src/arch/generic/decode_cache.cc
These classes are always used together, and merging them will give the ISAs
more flexibility in how they cache things and manage the process.
--HG--
rename : src/arch/x86/predecoder_tables.cc => src/arch/x86/decoder_tables.cc
This patch moves the ECF and EZF bits to individual registers (ecfBit and
ezfBit) and the CF and OF bits to cfofFlag registers. This is being done
so as to lower the read after write dependencies on the the condition code
register. Ultimately we will have the following registers [ZAPS], [OF],
[CF], [ECF], [EZF] and [DF]. Note that this is only one part of the
solution for lowering the dependencies. The other part will check whether
or not the condition code register needs to be actually read. This would
be done through a separate patch.
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.
It's possible for two page table walks to overlap which will go in the same
place in the TLB's trie. They would land on top of each other, so this change
adds some code which detects if an address already matches an entry and if so
throws away the new one.
The parameter is _machInst, which is very similar to the member machInst. If
machInst is used to pass the parameter to a lower level constructor, what
really happens is that machInst is set to whatever it already happened to be,
effectively leaving it uninitialized.
This change also adjusts the TlbEntry class so that it stores the number of
address bits wide a page is rather than its size in bytes. In other words,
instead of storing 4K for a 4K page, it stores 12. 12 is easy to turn into 4K,
but it's a little harder going the other way.
