Prior to this commit, all qstrs were required to be allocated (by calling
mp_emit_common_use_qstr) in the MP_PASS_SCOPE pass (the first one). But
this is an unnecessary restriction, which is lifted by this commit.
Lifting the restriction simplifies the compiler because it can allocate
qstrs in later passes.
This also generates better code, because in some cases (eg when a variable
is closed over) the scope of an identifier is not known until a bit later
and then the identifier no longer needs its qstr allocated in the global
table.
Code size is reduced for all ports with this commit.
Signed-off-by: Damien George <damien@micropython.org>
This commit adds support to the parser so that tuples which contain only
constant elements (bool, int, str, bytes, etc) are immediately converted to
a tuple object. This makes it more efficient to use tuples containing
constant data because they no longer need to be created at runtime by the
bytecode (or native code).
Furthermore, with this improvement constant tuples that are part of frozen
code are now able to be stored fully in ROM (this will be implemented in
later commits).
Code size is increased by about 400 bytes on Cortex-M4 platforms.
See related issue #722.
Signed-off-by: Damien George <damien@micropython.org>
This is a partial implementation of PEP 448 to allow multiple ** unpackings
when calling a function or method.
The compiler is modified to encode the argument as a None: obj key-value
pair (similar to how regular keyword arguments are encoded as str: obj
pairs). The extra object that was pushed on the stack to hold a single **
unpacking object is no longer used and is removed.
The runtime is modified to decode this new format.
Signed-off-by: David Lechner <david@pybricks.com>
This commit introduces changes:
- All jump opcodes are changed to have variable length arguments, of either
1 or 2 bytes (previously they were fixed at 2 bytes). In most cases only
1 byte is needed to encode the short jump offset, saving bytecode size.
- The bytecode emitter now selects 1 byte jump arguments when the jump
offset is guaranteed to fit in 1 byte. This is achieved by checking if
the code size changed during the last pass and, if it did (if it shrank),
then requesting that the compiler make another pass to get the correct
offsets of the now-smaller code. This can continue multiple times until
the code stabilises. The code can only ever shrink so this iteration is
guaranteed to complete. In most cases no extra passes are needed, the
original 4 passes are enough to get it right by the 4th pass (because the
2nd pass computes roughly the correct labels and the 3rd pass computes
the correct size for the jump argument).
This change to the jump opcode encoding reduces .mpy files and RAM usage
(when bytecode is in RAM) by about 2% on average.
The performance of the VM is not impacted, at least within measurment of
the performance benchmark suite.
Code size is reduced for builds that include a decent amount of frozen
bytecode. ARM Cortex-M builds without any frozen code increase by about
350 bytes.
Signed-off-by: Damien George <damien@micropython.org>
Background: .mpy files are precompiled .py files, built using mpy-cross,
that contain compiled bytecode functions (and can also contain machine
code). The benefit of using an .mpy file over a .py file is that they are
faster to import and take less memory when importing. They are also
smaller on disk.
But the real benefit of .mpy files comes when they are frozen into the
firmware. This is done by loading the .mpy file during compilation of the
firmware and turning it into a set of big C data structures (the job of
mpy-tool.py), which are then compiled and downloaded into the ROM of a
device. These C data structures can be executed in-place, ie directly from
ROM. This makes importing even faster because there is very little to do,
and also means such frozen modules take up much less RAM (because their
bytecode stays in ROM).
The downside of frozen code is that it requires recompiling and reflashing
the entire firmware. This can be a big barrier to entry, slows down
development time, and makes it harder to do OTA updates of frozen code
(because the whole firmware must be updated).
This commit attempts to solve this problem by providing a solution that
sits between loading .mpy files into RAM and freezing them into the
firmware. The .mpy file format has been reworked so that it consists of
data and bytecode which is mostly static and ready to run in-place. If
these new .mpy files are located in flash/ROM which is memory addressable,
the .mpy file can be executed (mostly) in-place.
With this approach there is still a small amount of unpacking and linking
of the .mpy file that needs to be done when it's imported, but it's still
much better than loading an .mpy from disk into RAM (although not as good
as freezing .mpy files into the firmware).
The main trick to make static .mpy files is to adjust the bytecode so any
qstrs that it references now go through a lookup table to convert from
local qstr number in the module to global qstr number in the firmware.
That means the bytecode does not need linking/rewriting of qstrs when it's
loaded. Instead only a small qstr table needs to be built (and put in RAM)
at import time. This means the bytecode itself is static/constant and can
be used directly if it's in addressable memory. Also the qstr string data
in the .mpy file, and some constant object data, can be used directly.
Note that the qstr table is global to the module (ie not per function).
In more detail, in the VM what used to be (schematically):
qst = DECODE_QSTR_VALUE;
is now (schematically):
idx = DECODE_QSTR_INDEX;
qst = qstr_table[idx];
That allows the bytecode to be fixed at compile time and not need
relinking/rewriting of the qstr values. Only qstr_table needs to be linked
when the .mpy is loaded.
Incidentally, this helps to reduce the size of bytecode because what used
to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices.
If the module uses the same qstr more than two times then the bytecode is
smaller than before.
The following changes are measured for this commit compared to the
previous (the baseline):
- average 7%-9% reduction in size of .mpy files
- frozen code size is reduced by about 5%-7%
- importing .py files uses about 5% less RAM in total
- importing .mpy files uses about 4% less RAM in total
- importing .py and .mpy files takes about the same time as before
The qstr indirection in the bytecode has only a small impact on VM
performance. For stm32 on PYBv1.0 the performance change of this commit
is:
diff of scores (higher is better)
N=100 M=100 baseline -> this-commit diff diff% (error%)
bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%)
bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%)
bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%)
bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%)
bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%)
bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%)
bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%)
core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%)
core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%)
core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%)
core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%)
misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%)
misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%)
misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%)
misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%)
viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%)
viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%)
viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%)
viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%)
viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%)
viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%)
And for unix on x64:
diff of scores (higher is better)
N=2000 M=2000 baseline -> this-commit diff diff% (error%)
bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%)
bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%)
bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%)
bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%)
bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%)
bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%)
bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%)
misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%)
misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%)
misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%)
misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%)
The code size change is (firmware with a lot of frozen code benefits the
most):
bare-arm: +396 +0.697%
minimal x86: +1595 +0.979% [incl +32(data)]
unix x64: +2408 +0.470% [incl +800(data)]
unix nanbox: +1396 +0.309% [incl -96(data)]
stm32: -1256 -0.318% PYBV10
cc3200: +288 +0.157%
esp8266: -260 -0.037% GENERIC
esp32: -216 -0.014% GENERIC[incl -1072(data)]
nrf: +116 +0.067% pca10040
rp2: -664 -0.135% PICO
samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS
As part of this change the .mpy file format version is bumped to version 6.
And mpy-tool.py has been improved to provide a good visualisation of the
contents of .mpy files.
In summary: this commit changes the bytecode to use qstr indirection, and
reworks the .mpy file format to be simpler and allow .mpy files to be
executed in-place. Performance is not impacted too much. Eventually it
will be possible to store such .mpy files in a linear, read-only, memory-
mappable filesystem so they can be executed from flash/ROM. This will
essentially be able to replace frozen code for most applications.
Signed-off-by: Damien George <damien@micropython.org>
This commit removes all parts of code associated with the existing
MICROPY_OPT_CACHE_MAP_LOOKUP_IN_BYTECODE optimisation option, including the
-mcache-lookup-bc option to mpy-cross.
This feature originally provided a significant performance boost for Unix,
but wasn't able to be enabled for MCU targets (due to frozen bytecode), and
added significant extra complexity to generating and distributing .mpy
files.
The equivalent performance gain is now provided by the combination of
MICROPY_OPT_LOAD_ATTR_FAST_PATH and MICROPY_OPT_MAP_LOOKUP_CACHE (which has
been enabled on the unix port in the previous commit).
It's hard to provide precise performance numbers, but tests have been run
on a wide variety of architectures (x86-64, ARM Cortex, Aarch64, RISC-V,
xtensa) and they all generally agree on the qualitative improvements seen
by the combination of MICROPY_OPT_LOAD_ATTR_FAST_PATH and
MICROPY_OPT_MAP_LOOKUP_CACHE.
For example, on a "quiet" Linux x64 environment (i3-5010U @ 2.10GHz) the
change from CACHE_MAP_LOOKUP_IN_BYTECODE, to LOAD_ATTR_FAST_PATH combined
with MAP_LOOKUP_CACHE is:
diff of scores (higher is better)
N=2000 M=2000 bccache -> attrmapcache diff diff% (error%)
bm_chaos.py 13742.56 -> 13905.67 : +163.11 = +1.187% (+/-3.75%)
bm_fannkuch.py 60.13 -> 61.34 : +1.21 = +2.012% (+/-2.11%)
bm_fft.py 113083.20 -> 114793.68 : +1710.48 = +1.513% (+/-1.57%)
bm_float.py 256552.80 -> 243908.29 : -12644.51 = -4.929% (+/-1.90%)
bm_hexiom.py 521.93 -> 625.41 : +103.48 = +19.826% (+/-0.40%)
bm_nqueens.py 197544.25 -> 217713.12 : +20168.87 = +10.210% (+/-3.01%)
bm_pidigits.py 8072.98 -> 8198.75 : +125.77 = +1.558% (+/-3.22%)
misc_aes.py 17283.45 -> 16480.52 : -802.93 = -4.646% (+/-0.82%)
misc_mandel.py 99083.99 -> 128939.84 : +29855.85 = +30.132% (+/-5.88%)
misc_pystone.py 83860.10 -> 82592.56 : -1267.54 = -1.511% (+/-2.27%)
misc_raytrace.py 21490.40 -> 22227.23 : +736.83 = +3.429% (+/-1.88%)
This shows that the new optimisations are at least as good as the existing
inline-bytecode-caching, and are sometimes much better (because the new
ones apply caching to a wider variety of map lookups).
The new optimisations can also benefit code generated by the native
emitter, because they apply to the runtime rather than the generated code.
The improvement for the native emitter when LOAD_ATTR_FAST_PATH and
MAP_LOOKUP_CACHE are enabled is (same Linux environment as above):
diff of scores (higher is better)
N=2000 M=2000 native -> nat-attrmapcache diff diff% (error%)
bm_chaos.py 14130.62 -> 15464.68 : +1334.06 = +9.441% (+/-7.11%)
bm_fannkuch.py 74.96 -> 76.16 : +1.20 = +1.601% (+/-1.80%)
bm_fft.py 166682.99 -> 168221.86 : +1538.87 = +0.923% (+/-4.20%)
bm_float.py 233415.23 -> 265524.90 : +32109.67 = +13.756% (+/-2.57%)
bm_hexiom.py 628.59 -> 734.17 : +105.58 = +16.796% (+/-1.39%)
bm_nqueens.py 225418.44 -> 232926.45 : +7508.01 = +3.331% (+/-3.10%)
bm_pidigits.py 6322.00 -> 6379.52 : +57.52 = +0.910% (+/-5.62%)
misc_aes.py 20670.10 -> 27223.18 : +6553.08 = +31.703% (+/-1.56%)
misc_mandel.py 138221.11 -> 152014.01 : +13792.90 = +9.979% (+/-2.46%)
misc_pystone.py 85032.14 -> 105681.44 : +20649.30 = +24.284% (+/-2.25%)
misc_raytrace.py 19800.01 -> 23350.73 : +3550.72 = +17.933% (+/-2.79%)
In summary, compared to MICROPY_OPT_CACHE_MAP_LOOKUP_IN_BYTECODE, the new
MICROPY_OPT_LOAD_ATTR_FAST_PATH and MICROPY_OPT_MAP_LOOKUP_CACHE options:
- are simpler;
- take less code size;
- are faster (generally);
- work with code generated by the native emitter;
- can be used on embedded targets with a small and constant RAM overhead;
- allow the same .mpy bytecode to run on all targets.
See #7680 for further discussion. And see also #7653 for a discussion
about simplifying mpy-cross options.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This implements (most of) the PEP-498 spec for f-strings and is based on
https://github.com/micropython/micropython/pull/4998 by @klardotsh.
It is implemented in the lexer as a syntax translation to `str.format`:
f"{a}" --> "{}".format(a)
It also supports:
f"{a=}" --> "a={}".format(a)
This is done by extracting the arguments into a temporary vstr buffer,
then after the string has been tokenized, the lexer input queue is saved
and the contents of the temporary vstr buffer are injected into the lexer
instead.
There are four main limitations:
- raw f-strings (`fr` or `rf` prefixes) are not supported and will raise
`SyntaxError: raw f-strings are not supported`.
- literal concatenation of f-strings with adjacent strings will fail
"{}" f"{a}" --> "{}{}".format(a) (str.format will incorrectly use
the braces from the non-f-string)
f"{a}" f"{a}" --> "{}".format(a) "{}".format(a) (cannot concatenate)
- PEP-498 requires the full parser to understand the interpolated
argument, however because this entirely runs in the lexer it cannot
resolve nested braces in expressions like
f"{'}'}"
- The !r, !s, and !a conversions are not supported.
Includes tests and cpydiffs.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Per CPython everything which comes after the command, module or file
argument is not an option for the interpreter itself. Hence the processing
of options should stop when encountering those, and the remainder be passed
as sys.argv. Note the latter was already the case for a module or file but
not for a command.
This fixes issues like 'micropython myfile.py -h' showing the help and
exiting instead of passing '-h' as sys.argv[1], likewise for
'-X <something>' being treated as a special option no matter where it
occurs on the command line.
The syntax matches CPython and the semantics are equivalent except that,
unlike CPython, MicroPython allows using := to assign to comprehension
iteration variables, because disallowing this would take a lot of code to
check for it.
The new compile-time option MICROPY_PY_ASSIGN_EXPR selects this feature and
is enabled by default, following MICROPY_PY_ASYNC_AWAIT.
This adds the Python files in the tests/ directory to be formatted with
./tools/codeformat.py. The basics/ subdirectory is excluded for now so we
aren't changing too much at once.
In a few places `# fmt: off`/`# fmt: on` was used where the code had
special formatting for readability or where the test was actually testing
the specific formatting.
When this variable is set to non-empty string it triggers the REPL after a
command/module/file finishes running.
The Python file without the file extension is because the cmdline: parser
in run-test splits on spaces, so we can't use the -c option since
`import os` can't be written without a space.
This commit adds backward-word, backward-kill-word, forward-word,
forward-kill-word sequences for the REPL, with bindings to Alt+F, Alt+B,
Alt+D and Alt+Backspace respectively. It is disabled by default and can be
enabled via MICROPY_REPL_EMACS_WORDS_MOVE.
Further enabling MICROPY_REPL_EMACS_EXTRA_WORDS_MOVE adds extra bindings
for these new sequences: Ctrl+Right, Ctrl+Left and Ctrl+W.
The features are enabled on unix micropython-coverage and micropython-dev.
This check follows CPython's behaviour, because 'import *' always populates
the globals with the imported names, not locals.
Since it's safe to do this (doesn't lead to a crash or undefined behaviour)
the check is only enabled for MICROPY_CPYTHON_COMPAT.
Fixes issue #5121.
This patch compresses the second part of the bytecode prelude which
contains the source file name, function name, source-line-number mapping
and cell closure information. This part of the prelude now begins with a
single varible length unsigned integer which encodes 2 numbers, being the
byte-size of the following 2 sections in the header: the "source info
section" and the "closure section". After decoding this variable unsigned
integer it's possible to skip over one or both of these sections very
easily.
This scheme saves about 2 bytes for most functions compared to the original
format: one in the case that there are no closure cells, and one because
padding was eliminated.
The start of the bytecode prelude contains 6 numbers telling the amount of
stack needed for the Python values and exceptions, and the signature of the
function. Prior to this patch these numbers were all encoded one after the
other (2x variable unsigned integers, then 4x bytes), but using so many
bytes is unnecessary.
An entropy analysis of around 150,000 bytecode functions from the CPython
standard library showed that the optimal Shannon coding would need about
7.1 bits on average to encode these 6 numbers, compared to the existing 48
bits.
This patch attempts to get close to this optimal value by packing the 6
numbers into a single, varible-length unsigned integer via bit-wise
interleaving. The interleaving scheme is chosen to minimise the average
number of bytes needed, and at the same time keep the scheme simple enough
so it can be implemented without too much overhead in code size or speed.
The scheme requires about 10.5 bits on average to store the 6 numbers.
As a result most functions which originally took 6 bytes to encode these 6
numbers now need only 1 byte (in 80% of cases).
From the beginning of this project the RAISE_VARARGS opcode was named and
implemented following CPython, where it has an argument (to the opcode)
counting how many args the raise takes:
raise # 0 args (re-raise previous exception)
raise exc # 1 arg
raise exc from exc2 # 2 args (chained raise)
In the bytecode this operation therefore takes 2 bytes, one for
RAISE_VARARGS and one for the number of args.
This patch splits this opcode into 3, where each is now a single byte.
This reduces bytecode size by 1 byte for each use of raise. Every byte
counts! It also has the benefit of reducing code size (on all ports except
nanbox).
To make progress towards MicroPython supporting Python 3.5, adding the
matmul operator is important because it's a really "low level" part of the
language, being a new token and modifications to the grammar.
It doesn't make sense to make it configurable because 1) it would make the
grammar and lexer complicated/messy; 2) no other operators are
configurable; 3) it's not a feature that can be "dynamically plugged in"
via an import.
And matmul can be useful as a general purpose user-defined operator, it
doesn't have to be just for numpy use.
Based on work done by Jim Mussared.
POP_BLOCK and POP_EXCEPT are now the same, and are always followed by a
JUMP. So this optimisation reduces code size, and RAM usage of bytecode by
two bytes for each try-except handler.
This patch fixes a bug in the VM when breaking within a try-finally. The
bug has to do with executing a break within the finally block of a
try-finally statement. For example:
def f():
for x in (1,):
print('a', x)
try:
raise Exception
finally:
print(1)
break
print('b', x)
f()
Currently in uPy the above code will print:
a 1
1
1
segmentation fault (core dumped) micropython
Not only is there a seg fault, but the "1" in the finally block is printed
twice. This is because when the VM executes a finally block it doesn't
really know if that block was executed due to a fall-through of the try (no
exception raised), or because an exception is active. In particular, for
nested finallys the VM has no idea which of the nested ones have active
exceptions and which are just fall-throughs. So when a break (or continue)
is executed it tries to unwind all of the finallys, when in fact only some
may be active.
It's questionable whether break (or return or continue) should be allowed
within a finally block, because they implicitly swallow any active
exception, but nevertheless it's allowed by CPython (although almost never
used in the standard library). And uPy should at least not crash in such a
case.
The solution here relies on the fact that exception and finally handlers
always appear in the bytecode after the try body.
Note: there was a similar bug with a return in a finally block, but that
was previously fixed in b735208403
The way it was written previously the variable x was not an implicit
nonlocal, it was just a normal local (but the compiler has a bug which
incorrectly makes it a nonlocal).
This patch changes the way REPL autocomplete finds matches. It now probes
the target object for all qstrs via mp_load_method_maybe to look for a
match with the given input string. Similar to how the builtin dir()
function works, this new algorithm now find all methods and instances of
user-defined classes including attributes of their parent classes. This
helps a lot at the REPL prompt for user-discovery and to autocomplete names
even for classes that are derived.
The downside is that this new algorithm is slower than the previous one,
and in particular will be slower the more qstrs there are in the system.
But because REPL autocomplete is primarily used in an interactive way it is
not that important to make it fast, as long as it is "fast enough" compared
to human reaction.
On a slow microcontroller (CPU running at 16MHz) the autocomplete time for
a list of 35 names in the outer namespace (pressing tab at a bare prompt)
takes about 160ms with this algorithm, compared to about 40ms for the
previous implementation (this time includes the actual printing of the
names as well). This time of 160ms is very reasonable especially given the
new functionality of listing all the names.
This patch also decreases code size by:
bare-arm: +0
minimal x86: -128
unix x64: -128
unix nanbox: -224
stm32: -88
cc3200: -80
esp8266: -92
esp32: -84
This is to allow to place reverse ops immediately after normal ops, so
they can be tested as one range (which is optimization for reverse ops
introduction in the next patch).
It starts a dichotomy of mp_binary_op_t values which can't appear in the
bytecode. Another reason to move it is to VALUES of OP_* and OP_INPLACE_*
nicely adjacent. This also will be needed for OP_REVERSE_*, to be soon
introduced.
This patch refactors the handling of the special super() call within the
compiler. It removes the need for a global (to the compiler) state variable
which keeps track of whether the subject of an expression is super. The
handling of super() is now done entirely within one function, which makes
the compiler a bit cleaner and allows to easily add more optimisations to
super calls.
Changes to the code size are:
bare-arm: +12
minimal: +0
unix x64: +48
unix nanbox: -16
stmhal: +4
cc3200: +0
esp8266: -56
Previous to this patch any non-interned str/bytes objects would create a
special parse node that held a copy of the str/bytes data. Then in the
compiler this data would be turned into a str/bytes object. This actually
lead to 2 copies of the data, one in the parse node and one in the object.
The parse node's copy of the data would be freed at the end of the compile
stage but nevertheless it meant that the peak memory usage of the
parse/compile stage was higher than it needed to be (by an amount equal to
the number of bytes in all the non-interned str/bytes objects).
This patch changes the behaviour so that str/bytes objects are created
directly in the parser and the object stored in a const-object parse node
(which already exists for bignum, float and complex const objects). This
reduces peak RAM usage of the parse/compile stage, simplifies the parser
and compiler, and reduces code size by about 170 bytes on Thumb2 archs,
and by about 300 bytes on Xtensa archs.
Dan Halbert
Damien George
David Lechner
David Lechner
Jim Mussared
stijn
Yonatan Goldschmidt
Petr Viktorin
Damien George
Milan Rossa
Paul Sokolovsky
Rami Ali