We currently generate and compile one version of the ISA code per CPU
model. This is obviously wasting a lot of resources at compile
time. This changeset factors out the interface into a separate
ExecContext class, which also serves as documentation for the
interface between CPUs and the ISA code. While doing so, this
changeset also fixes up interface inconsistencies between the
different CPU models.
The main argument for using one set of ISA code per CPU model has
always been performance as this avoid indirect branches in the
generated code. However, this argument does not hold water. Booting
Linux on a simulated ARM system running in atomic mode
(opt/10.linux-boot/realview-simple-atomic) is actually 2% faster
(compiled using clang 3.4) after applying this patch. Additionally,
compilation time is decreased by 35%.
This changeset fixes three types of warnings that occur in clang 3.4
on Ubuntu 12.04:
* Certain versions of libstdc++ (primarily 4.8) use struct and class
interchangeably. This triggers a warning in clang.
* Swig has a tendency to generate code with the register class which
was deprecated in C++11. This triggers a deprecation warning in
clang.
* Swig sometimes generates Python wrapper code which returns
uninitialized values. It's unclear if this is actually a problem
(the cases might be limited to failure paths). We'll silence these
warnings for now since there is little we can do about the
generated code.
It seems gcc >4.8 does not get along well with binutils <= 2.22, and
to help users this patch adds a warning with an indication for how to
fix the issue. It might even be worth adding a Exit(-1) and stop the
build.
This patch bumps the supported version of gcc from 4.4 to 4.6, and
clang from 2.9 to 3.0. This enables, amongst other things, range-based
for loops, lambda expressions, etc. The STL implementation shipping
with 4.6 also has a full functional implementation of unique_ptr and
shared_ptr.
This patch encompasses several interrelated and interdependent changes
to the ISA generation step. The end goal is to reduce the size of the
generated compilation units for instruction execution and decoding so
that batch compilation can proceed with all CPUs active without
exhausting physical memory.
The ISA parser (src/arch/isa_parser.py) has been improved so that it can
accept 'split [output_type];' directives at the top level of the grammar
and 'split(output_type)' python calls within 'exec {{ ... }}' blocks.
This has the effect of "splitting" the files into smaller compilation
units. I use air-quotes around "splitting" because the files themselves
are not split, but preprocessing directives are inserted to have the same
effect.
Architecturally, the ISA parser has had some changes in how it works.
In general, it emits code sooner. It doesn't generate per-CPU files,
and instead defers to the C preprocessor to create the duplicate copies
for each CPU type. Likewise there are more files emitted and the C
preprocessor does more substitution that used to be done by the ISA parser.
Finally, the build system (SCons) needs to be able to cope with a
dynamic list of source files coming out of the ISA parser. The changes
to the SCons{cript,truct} files support this. In broad strokes, the
targets requested on the command line are hidden from SCons until all
the build dependencies are determined, otherwise it would try, realize
it can't reach the goal, and terminate in failure. Since build steps
(i.e. running the ISA parser) must be taken to determine the file list,
several new build stages have been inserted at the very start of the
build. First, the build dependencies from the ISA parser will be emitted
to arch/$ISA/generated/inc.d, which is then read by a new SCons builder
to finalize the dependencies. (Once inc.d exists, the ISA parser will not
need to be run to complete this step.) Once the dependencies are known,
the 'Environments' are made by the makeEnv() function. This function used
to be called before the build began but now happens during the build.
It is easy to see that this step is quite slow; this is a known issue
and it's important to realize that it was already slow, but there was
no obvious cause to attribute it to since nothing was displayed to the
terminal. Since new steps that used to be performed serially are now in a
potentially-parallel build phase, the pathname handling in the SCons scripts
has been tightened up to deal with chdir() race conditions. In general,
pathnames are computed earlier and more likely to be stored, passed around,
and processed as absolute paths rather than relative paths. In the end,
some of these issues had to be fixed by inserting serializing dependencies
in the build.
Minor note:
For the null ISA, we just provide a dummy inc.d so SCons is never
compelled to try to generate it. While it seems slightly wrong to have
anything in src/arch/*/generated (i.e. a non-generated 'generated' file),
it's by far the simplest solution.
This patch fixes an issue with the way the python-config path is
parsed, as it caused issues on systems where a newline ended up being
included in the path.
Compile gem5 on systems where python2 and python3 co-exists without any
changes in path. python2-config is chosen over python-config if it exists.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
This patch adds DRAMSim2 as a memory controller by wrapping the
external library and creating a sublass of AbstractMemory that bridges
between the semantics of gem5 and the DRAMSim2 interface.
The DRAMSim2 wrapper extracts the clock period from the config
file. There is no way of extracting this information from DRAMSim2
itself, so we simply read the same config file and get it from there.
To properly model the response queue, the wrapper keeps track of how
many transactions are in the actual controller, and how many are
stacking up waiting to be sent back as responses (in the wrapper). The
latter requires us to move away from the queued port and manage the
packets ourselves. This is due to DRAMSim2 not having any flow control
on the response path.
DRAMSim2 assumes that the transactions it is given are matching the
burst size of the choosen memory. The wrapper checks to ensure the
cache line size of the system matches the burst size of DRAMSim2 as
there are currently no provisions to split the system requests. In
theory we could allow a cache line size smaller than the burst size,
but that would lead to inefficient use of the DRAM, so for not we
fatal also in this case.
The performance counting framework in Linux 3.2 and onwards supports
an attribute to exclude events generated by the host when running
KVM. Setting this attribute allows us to get more reliable
measurements of the guest machine. For example, on a highly loaded
system, the instruction counts from the guest can be severely
distorted by the host kernel (e.g., by page fault handlers).
This changeset introduces a check for the attribute and enables it in
the KVM CPU if present.
The kvm-based CPU module requires support for POSIX timers. This
changeset adds a check for POSIX timers and ensures that gem5 is
linked with librt if necessary. KVM support is disabled if POSIX
timers are not supported by the host. This fixes a compilation issue
for some glibc versions where clock_nanosleep and timer_create are in
different libraries.
This changeset includes libfputils from revision bbf0d61d75. This
library can be used to convert to and from 80-bit floats and query the
type of an 80-bit float, which is needed to support the x87 FPU.
This changeset adds support for KVM on x86. Full support is split
across a number of commits since some features are relatively
complex. This changeset includes support for:
* Integer state synchronization (including segment regs)
* CPUID (gem5's CPUID values are inserted into KVM)
* x86 legacy IO (remapped and handled by gem5's memory system)
* Memory mapped IO
* PCI
* MSRs
* State dumping
Most of the functionality is fairly straight forward. There are some
quirks to support PCI enumerations since this is done in the TLB(!) in
the simulated CPUs. We currently replicate some of that code.
Unlike the ARM implementation, the x86 implementation of the virtual
CPU does not use the cycles hardware counter. KVM on x86 simulates the
time stamp counter (TSC) in the kernel. If we just measure host cycles
using perfevent, we might end up measuring a slightly different number
of cycles. If we don't get the cycle accounting right, we might end up
rewinding the TSC, with all kinds of chaos as a result.
An additional feature of the KVM CPU on x86 is extended state
dumping. This enables Python scripts controlling the simulator to
request dumping of a subset of the processor state. The following
methods are currenlty supported:
* dumpFpuRegs
* dumpIntRegs
* dumpSpecRegs
* dumpDebugRegs
* dumpXCRs
* dumpXSave
* dumpVCpuEvents
* dumpMSRs
Known limitations:
* M5 ops are currently not supported.
* FPU synchronization is not supported (only affects CPU switching).
Both of the limitations will be addressed in separate commits.
This patch changes the SConscript to build gem5 with libc++ on OSX as
the conventional libstdc++ does not have the C++11 constructs that the
current code base makes use of (e.g. std::forward).
Since this was the last use of the transitional TR1, the unordered map
and set header can now be simplified as well.
This patch changes how we determine the Python-related compiler and
linker flags. The previous approach used the internal LINKFORSHARED
which is not intended as part of the external API
(http://bugs.python.org/issue3588) and causes failures on recent OSX
installations.
Instead of using distutils we now rely on python-config and scons
ParseConfig. For backwards compatibility we also parse out the
includes and libs although this could safely be dropped. The drawback
of this patch is that Python 2.5 is now required, but hopefully that
is an acceptable compromise as any system with gcc 4.4 most likely
will have Python >= 2.5.
Architecture specific limitations:
* LPAE is currently not supported by gem5. We therefore panic if LPAE
is enabled when returning to gem5.
* The co-processor based interface to the architected timer is
unsupported. We can't support this due to limitations in the KVM
API on ARM.
* M5 ops are currently not supported. This requires either a kernel
hack or a memory mapped device that handles the guest<->m5
interface.
This changeset introduces the architecture independent parts required
to support KVM-accelerated CPUs. It introduces two new simulation
objects:
KvmVM -- The KVM VM is a component shared between all CPUs in a shared
memory domain. It is typically instantiated as a child of the
system object in the simulation hierarchy. It provides access
to KVM VM specific interfaces.
BaseKvmCPU -- Abstract base class for all KVM-based CPUs. Architecture
dependent CPU implementations inherit from this class
and implement the following methods:
* updateKvmState() -- Update the
architecture-dependent KVM state from the gem5
thread context associated with the CPU.
* updateThreadContext() -- Update the thread context
from the architecture-dependent KVM state.
* dump() -- Dump the KVM state using (optional).
In order to deliver interrupts to the guest, CPU
implementations typically override the tick() method and
check for, and deliver, interrupts prior to entering
KVM.
Hardware-virutalized CPU currently have the following limitations:
* SE mode is not supported.
* PC events are not supported.
* Timing statistics are currently very limited. The current approach
simply scales the host cycles with a user-configurable factor.
* The simulated system must not contain any caches.
* Since cycle counts are approximate, there is no way to request an
exact number of cycles (or instructions) to be executed by the CPU.
* Hardware virtualized CPUs and gem5 CPUs must not execute at the
same time in the same simulator instance.
* Only single-CPU systems can be simulated.
* Remote GDB connections to the guest system are not supported.
Additionally, m5ops requires an architecture specific interface and
might not be supported.
tcmalloc_minimal doesn't support the heap checker on Debian, while
tcmalloc does. Instead of always linking with tcmalloc_minimal, if it
exists, we first check for tcmalloc and then use tcmalloc_minimal as a
fallback.
According to the tcmalloc readme, the recommended way of compiling
applications that make use of tcmalloc is to disable compiler
optimizations that make assumptions about malloc and friends. This
changeset adds the necessary compiler flags for both gcc and clang.
From the tcmalloc readme:
"NOTE: When compiling with programs with gcc, that you plan to link
with libtcmalloc, it's safest to pass in the flags
-fno-builtin-malloc -fno-builtin-calloc
-fno-builtin-realloc -fno-builtin-free
when compiling."
Python requires the flags in LINKFORSHARED to be added the linker
flags when linking with a statically with Python. Failing to do so can
lead to errors from the Python's dynamic module loader at start up.
--HG--
extra : rebase_source : e7a8daf72f4ede7ee5a4a5398a0b12e978a919b9
SWIG version 2.0.9 uses fully qualified module names despite of the
importing module being in the same package as the imported
module. This has the unfortunate consequence of causing the following
error when importing m5.internal.event:
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "src/python/importer.py", line 75, in load_module
exec code in mod.__dict__
File "src/python/m5/__init__.py", line 35, in <module>
import internal
File "src/python/importer.py", line 75, in load_module
exec code in mod.__dict__
File "src/python/m5/internal/__init__.py", line 32, in <module>
import event
File "src/python/importer.py", line 75, in load_module
exec code in mod.__dict__
File "build/X86/python/swig/event.py", line 107, in <module>
class Event(m5.internal.serialize.Serializable):
AttributeError: 'module' object has no attribute 'internal'
When 'event' is loaded, it triggers 'serialize' to be loaded. However,
it seems like the dictionary of 'm5' isn't updated until after
__init__.py terminates, which means that 'event' never sees the
'internal' attribute on 'm5'. Older versions of SWIG didn't include
the fully qualified module name if the modules were in the same
package.
This patch restructures and unifies the flags used by gcc and clang as
they are largely the same. The common parts are now dealt with in a
shared block of code, and the few bits and pieces that are
specifically affecting either gcc or clang are done separately.
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 adds libfdt, a library necessary for supporting
flattened device tree support in current and future versions of
the linux/android kernel for ARM.
This patch changes the use of pkg-config such that protobuf is still
evaluated with CheckLib even if it fails. This is to allow setups
where libprotobuf is available, but not configured through
protobuf. Moreover, if CheckLib fails to use libprotobuf then all the
tracing is disabled, but scons is allowed to continue with a warning.
This patch checks that the compiler in use is either gcc >= 4.4 or
clang >= 2.9. and enables building with --std=c++0x in all cases. As a
consequence, we can tidy up the hashmap and always have static_assert
available. If anyone wants to use alternative compilers, icc for
example supports c++0x to a similar level and could be added if
needed.
This patch opens up for a more elaborate use of c++0x features that
are present in gcc 4.4 and clang 2.9, e.g. auto typed variables,
variadic templates, rvalues and move semantics, and strongly typed
enums. There will be no going back on this one...
This patch simply prunes the SUNCC and ICC compiler options as they
are both sufficiently stale that they would have to be re-written from
scratch anyhow. The patch serves to clean things up before shifting to
a build environment that enforces basic c++11 compliance as done in
the following patch.
Scons normally removes all environment variables that aren't
whitelisted from the build environment. This messes up things like
ccache, distcc, and the clang static analyzer. This changeset adds the
DISTCC_, CCACHE_, and CCC_ prefixes to the environment variable
whitelist.
This patch enables the use of protobuf input files in the build
process, thus allowing .proto files to be added to input. Each .proto
file is compiled using the protoc tool and the newly created C++
source is added to the list of sources.
The first location where the protobufs will be used is in the
capturing and replay of memory traces, involving the communication
monitor and the trace-generator state of the traffic generator. This
will follow in the next patch.
This patch does add a dependency on the availability of the BSD
licensed protobuf library (and headers), and the protobuf compiler,
protoc. These dependencies are checked in the SConstruct, similar to
e.g. swig. The user can override the use of protoc from the PATH by
specifying the PROTOC environment variable.
Although the dependency on libprotobuf and protoc might seem like a
big step, they add significant value to the project going
forward. Execution traces and other types of traces could easily be
added and parsers for C++ and Python are automatically generated. We
could also envision using protobufs for the checkpoints, description
of the traffic-generator behaviour etc. The sky is the limit. We could
also use the GzipOutputStream from the protobuf library instead of the
current GPL gzstream.
Currently, only the C++ source and header is generated. Going forward
we might want to add the Python output to support simple command-line
tools for displaying and editing the traces.
C++11 has support for static_asserts to provide compile-time assertion
checking. This is very useful when testing, for example, structure
sizes to make sure that the compiler got the right alignment or vector
sizes.
This patch adds a few more checks to ensure that a compiler is present
on the system, along with swig. It references the relevant packages on
Ubuntu/RedHat, and also adds a similar line for the Python headers.
This patch adds Link-Time Optimization when building the fast target
using gcc >= 4.6, and adds a scons flag to disable it (-no-lto). No
check is performed to guarantee that the linker supports LTO and use
of the linker plugin, so the user has to ensure that binutils GNU ld
>= 2.21 or the gold linker is available. Typically, if gcc >= 4.6 is
available, the latter should not be a problem. Currently the LTO
option is only useful for gcc >= 4.6, due to the limited support on
clang and earlier versions of gcc. The intention is to also add
support for clang once the LTO integration matures.
The same number of jobs is used for the parallel phase of LTO as the
jobs specified on the scons command line, using the -flto=n flag that
was introduced with gcc 4.6. The gold linker also supports concurrent
and incremental linking, but this is not used at this point.
The compilation and linking time is increased by almost 50% on
average, although ARM seems to be particularly demanding with an
increase of almost 100%. Also beware when using this as gcc uses a
tremendous amount of memory and temp space in the process. You have
been warned.
After some careful consideration, and plenty discussions, the flag is
only added to the fast target, and the warning that was issued in an
earlier version of this patch is now removed. Similarly, the flag used
to enable LTO, now the default is to use it, and the flag has been
modified to disable LTO. The rationale behind this decision is that
opt is used for development, whereas fast is only used for long runs,
e.g. regressions or more elaborate experiments where the additional
compile and link time is amortized by a much larger run time.
When it comes to the return on investment, the regression seems to be
roughly 15% faster with LTO. For a bit more detail, I ran twolf on
ARM.fast, with three repeated runs, and they all finish within 42
minutes (+- 25 seconds) without LTO and 31 minutes (+- 25 seconds)
with LTO, i.e. LTO gives an impressive >25% speed-up for this case.
Without LTO (ARM.fast twolf)
real 42m37.632s
user 42m34.448s
sys 0m0.390s
real 41m51.793s
user 41m50.384s
sys 0m0.131s
real 41m45.491s
user 41m39.791s
sys 0m0.139s
With LTO (ARM.fast twolf)
real 30m33.588s
user 30m5.701s
sys 0m0.141s
real 31m27.791s
user 31m24.674s
sys 0m0.111s
real 31m25.500s
user 31m16.731s
sys 0m0.106s
This patch shifts the version of gcc for which we enable c++0x from
4.6 to 4.4 The more long term plan is to see what the c++0x features
can bring and what level of support would be enabled simply by bumping
the required version of gcc from 4.3 to 4.4.
A few minor things had to be fixed in the code base, most notably the
choice of a hashmap implementation. In the Ruby Sequencer there were
also a few minor issues that gcc 4.4 was not too happy about.
This patch adds the LIBRARY_PATH from the users OS environment to
Scons build environment. This path is used when linking to search for
libraries, and this patch enables tcmalloc to be used during the build
even if it is not placed in the default search paths.
This patch updates the message printed if the user does not have
tcmalloc available. It turns out that the correct package (which
creates all required symlinks etc) is libgoogle-perftools-dev. This
has been verified on Ubuntu 12.04.
This package is available in Ubuntu, Debian, and Redhat as google-perftools.
With multiple tests on a single machine I've seen a little over 10% performance
gain with tcmalloc.
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.
