This patch is a code optimisation, trading text bytes for speed. On
pyboard it's an increase of 0.06% in code size for a gain (in pystone
performance) of roughly 6.5%.
The patch optimises load/store/delete of attributes in user defined classes
by not looking up special accessors (@property, __get__, __delete__,
__set__, __setattr__ and __getattr_) if they are guaranteed not to exist in
the class.
Currently, if you do my_obj.foo() then the runtime has to do a few checks
to see if foo is a property or has __get__, and if so delegate the call.
And for stores things like my_obj.foo = 1 has to first check if foo is a
property or has __set__ defined on it.
Doing all those checks each and every time the attribute is accessed has a
performance penalty. This patch eliminates all those checks for cases when
it's guaranteed that the checks will always fail, ie no attributes are
properties nor have any special accessor methods defined on them.
To make this guarantee it checks all attributes of a user-defined class
when it is first created. If any of the attributes of the user class are
properties or have special accessors, or any of the base classes of the
user class have them, then it sets a flag in the class to indicate that
special accessors must be checked for. Then in the load/store/delete code
it checks this flag to see if it can take the shortcut and optimise the
lookup.
It's an optimisation that's pretty widely applicable because it improves
lookup performance for all methods of user defined classes, and stores of
attributes, at least for those that don't have special accessors. And, it
allows to enable descriptors with minimal additional runtime overhead if
they are not used for a particular user class.
There is one restriction on dynamic class creation that has been introduced
by this patch: a user-defined class cannot go from zero special accessors
to one special accessor (or more) after that class has been subclassed. If
the script attempts this an AttributeError is raised (see addition to
tests/misc/non_compliant.py for an example of this case).
The cost in code space bytes for the optimisation in this patch is:
unix x64: +528
unix nanbox: +508
stm32: +192
cc3200: +200
esp8266: +332
esp32: +244
Performance tests that were done:
- on unix x86-64, pystone improved by about 5%
- on pyboard, pystone improved by about 6.5%, from 1683 up to 1794
- on pyboard, bm_chaos (from CPython benchmark suite) improved by about 5%
- on esp32, pystone improved by about 30% (but there are caching effects)
- on esp32, bm_chaos improved by about 11%
Note that the check for elem!=NULL is removed for the
MP_MAP_LOOKUP_ADD_IF_NOT_FOUND case because mp_map_lookup will always
return non-NULL for such a case.
Before this patch, if a user defined the __new__() function for a class
then two instances of that class would be created: once before __new__ is
called and once during the __new__ call (assuming the user creates some
instance, eg using super().__new__, which is most of the time). The first
one was then discarded. This refactor makes it so that a new instance is
only created if the user __new__ function doesn't exist.
This patch cleans up and generalises part of the code which handles
overriding and calling a native base-class's __init__ method. It defers
the call to the native make_new() function until after the user (Python)
__init__() method has run. That user method now has the chance to call the
native __init__/make_new and pass it different arguments. If the user
doesn't call the super().__init__ method then it will be called
automatically after the user code finishes, to finalise construction of the
instance.
Before this patch MP_BINARY_OP_IN had two meanings: coming from bytecode it
meant that the args needed to be swapped, but coming from within the
runtime meant that the args were already in the correct order. This lead
to some confusion in the code and comments stating how args were reversed.
It also lead to 2 bugs: 1) containment for a subclass of a native type
didn't work; 2) the expression "{True} in True" would illegally succeed and
return True. In both of these cases it was because the args to
MP_BINARY_OP_IN ended up being reversed twice.
To fix these things this patch introduces MP_BINARY_OP_CONTAINS which
corresponds exactly to the __contains__ special method, and this is the
operator that built-in types should implement. MP_BINARY_OP_IN is now only
emitted by the compiler and is converted to MP_BINARY_OP_CONTAINS by
swapping the arguments.
This patch introduces a new compile-time config option to disable multiple
inheritance at the Python level: MICROPY_MULTIPLE_INHERITANCE. It is
enabled by default.
Disabling multiple inheritance eliminates a lot of recursion in the call
graph (which is important for some embedded systems), and can be used to
reduce code size for ports that are really constrained (by around 200 bytes
for Thumb2 archs).
With multiple inheritance disabled all tests in the test-suite pass except
those that explicitly test for multiple inheritance.
This allows to configure support for inplace special methods separately,
similar to "normal" and reverse special methods. This is useful, because
inplace methods are "the most optional" ones, for example, if inplace
methods aren't defined, the operation will be executed using normal
methods instead.
As a caveat, __iadd__ and __isub__ are implemented even if
MICROPY_PY_ALL_INPLACE_SPECIAL_METHODS isn't defined. This is similar
to the state of affairs before binary operations refactor, and allows
to run existing tests even if MICROPY_PY_ALL_INPLACE_SPECIAL_METHODS
isn't defined.
If MICROPY_PY_ALL_SPECIAL_METHODS is defined, actually define all special
methods (still subject to gating by e.g. MICROPY_PY_REVERSE_SPECIAL_METHODS).
This adds quite a number of qstr's, so should be used sparingly.
Update makeqstrdata.py to sort strings starting with "__" to the beginning
of qstr list, so they get low qstr id's, guaranteedly fitting in 8 bits.
Then use this property to further compact op_id => qstr mapping arrays.
Per https://docs.python.org/3/library/sys.html#sys.getsizeof:
getsizeof() calls the object’s __sizeof__ method. Previously, "getsizeof"
was used mostly to save on new qstr, as we don't really support calling
this method on arbitrary objects (so it was used only for reporting).
However, normalize it all now.
Header files that are considered internal to the py core and should not
normally be included directly are:
py/nlr.h - internal nlr configuration and declarations
py/bc0.h - contains bytecode macro definitions
py/runtime0.h - contains basic runtime enums
Instead, the top-level header files to include are one of:
py/obj.h - includes runtime0.h and defines everything to use the
mp_obj_t type
py/runtime.h - includes mpstate.h and hence nlr.h, obj.h, runtime0.h,
and defines everything to use the general runtime support functions
Additional, specific headers (eg py/objlist.h) can be included if needed.
Qstr values fit in 16-bits (and this fact is used elsewhere in the code) so
no need to use more than that for the large lookup tables. The compiler
will anyway give a warning if the qstr values don't fit in 16 bits. Saves
around 80 bytes of code space for Thumb2 archs.
If, for class X, X.__add__(Y) doesn't exist (or returns NotImplemented),
try Y.__radd__(X) instead.
This patch could be simpler, but requires undoing operand swap and
operation switch to get non-confusing error message in case __radd__
doesn't exist.
NotImplemented means "try other fallbacks (like calling __rop__
instead of __op__) and if nothing works, raise TypeError". As
MicroPython doesn't implement any fallbacks, signal to raise
TypeError right away.
The unary-op/binary-op enums are already defined, and there are no
arithmetic tricks used with these types, so it makes sense to use the
correct enum type for arguments that take these values. It also reduces
code size quite a bit for nan-boxing builds.
This patch allows the following code to run without allocating on the heap:
super().foo(...)
Before this patch such a call would allocate a super object on the heap and
then load the foo method and call it right away. The super object is only
needed to perform the lookup of the method and not needed after that. This
patch makes an optimisation to allocate the super object on the C stack and
discard it right after use.
Changes in code size due to this patch are:
bare-arm: +128
minimal: +232
unix x64: +416
unix nanbox: +364
stmhal: +184
esp8266: +340
cc3200: +128
The common cases for inheritance are 0 or 1 parent types, for both built-in
types (eg built-in exceptions) as well as user defined types. So it makes
sense to optimise the case of 1 parent type by storing just the type and
not a tuple of 1 value (that value being the single parent type).
This patch makes such an optimisation. Even though there is a bit more
code to handle the two cases (either a single type or a tuple with 2 or
more values) it helps reduce overall code size because it eliminates the
need to create a static tuple to hold single parents (eg for the built-in
exceptions). It also helps reduce RAM usage for user defined types that
only derive from a single parent.
Changes in code size (in bytes) due to this patch:
bare-arm: -16
minimal (x86): -176
unix (x86-64): -320
unix nanbox: -384
stmhal: -64
cc3200: -32
esp8266: -108
This patch changes mp_uint_t to size_t for the len argument of the
following public facing C functions:
mp_obj_tuple_get
mp_obj_list_get
mp_obj_get_array
These functions take a pointer to the len argument (to be filled in by the
function) and callers of these functions should update their code so the
type of len is changed to size_t. For ports that don't use nan-boxing
there should be no change in generate code because the size of the type
remains the same (word sized), and in a lot of cases there won't even be a
compiler warning if the type remains as mp_uint_t.
The reason for this change is to standardise on the use of size_t for
variables that count memory (or memory related) sizes/lengths. It helps
builds that use nan-boxing.
Allows to iterate over the following without allocating on the heap:
- tuple
- list
- string, bytes
- bytearray, array
- dict (not dict.keys, dict.values, dict.items)
- set, frozenset
Allows to call the following without heap memory:
- all, any, min, max, sum
TODO: still need to allocate stack memory in bytecode for iter_buf.
This patch implements support for class methods __delattr__ and __setattr__
for customising attribute access. It is controlled by the config option
MICROPY_PY_DELATTR_SETATTR and is disabled by default.
This allows to define an abstract base class which would translate
C-level protocol to Python method calls, and any subclass inheriting
from it will support this feature. This in particular actually enables
recently introduced machine.PinBase class.
The first argument to the type.make_new method is naturally a uPy type,
and all uses of this argument cast it directly to a pointer to a type
structure. So it makes sense to just have it a pointer to a type from
the very beginning (and a const pointer at that). This patch makes
such a change, and removes all unnecessary casting to/from mp_obj_t.
This patch changes the type signature of .make_new and .call object method
slots to use size_t for n_args and n_kw (was mp_uint_t. Makes code more
efficient when mp_uint_t is larger than a machine word. Doesn't affect
ports when size_t and mp_uint_t have the same size.
Note that even though wrapped in MICROPY_CPYTHON_COMPAT, it is not
fully compatible because the modifications to the dictionary do not
propagate to the actual instance members.
This allows the mp_obj_t type to be configured to something other than a
pointer-sized primitive type.
This patch also includes additional changes to allow the code to compile
when sizeof(mp_uint_t) != sizeof(void*), such as using size_t instead of
mp_uint_t, and various casts.
mp_obj_get_int_truncated will raise a TypeError if the argument is not
an integral type. Use mp_obj_int_get_truncated only when you know the
argument is a small or big int.
Hashing is now done using mp_unary_op function with MP_UNARY_OP_HASH as
the operator argument. Hashing for int, str and bytes still go via
fast-path in mp_unary_op since they are the most common objects which
need to be hashed.
This lead to quite a bit of code cleanup, and should be more efficient
if anything. It saves 176 bytes code space on Thumb2, and 360 bytes on
x86.
The only loss is that the error message "unhashable type" is now the
more generic "unsupported type for __hash__".
Previous to this patch the printing mechanism was a bit of a tangled
mess. This patch attempts to consolidate printing into one interface.
All (non-debug) printing now uses the mp_print* family of functions,
mainly mp_printf. All these functions take an mp_print_t structure as
their first argument, and this structure defines the printing backend
through the "print_strn" function of said structure.
Printing from the uPy core can reach the platform-defined print code via
two paths: either through mp_sys_stdout_obj (defined pert port) in
conjunction with mp_stream_write; or through the mp_plat_print structure
which uses the MP_PLAT_PRINT_STRN macro to define how string are printed
on the platform. The former is only used when MICROPY_PY_IO is defined.
With this new scheme printing is generally more efficient (less layers
to go through, less arguments to pass), and, given an mp_print_t*
structure, one can call mp_print_str for efficiency instead of
mp_printf("%s", ...). Code size is also reduced by around 200 bytes on
Thumb2 archs.
This simplifies the API for objects and reduces code size (by around 400
bytes on Thumb2, and around 2k on x86). Performance impact was measured
with Pystone score, but change was barely noticeable.
Despite initial guess, this code factoring does not hamper performance.
In fact it seems to improve speed by a little: running pystone(1.2) on
pyboard (which gives a very stable result) this patch takes pystones
from 1729.51 up to 1742.16. Also, pystones on x64 increase by around
the same proportion (but it's much noisier).
Taking a look at the generated machine code, stack usage with this patch
is unchanged, and call is tail-optimised with all arguments in
registers. Code size decreases by about 50 bytes on Thumb2 archs.
Eg, "() + 1" now tells you that __add__ is not supported for tuple and
int types (before it just said the generic "binary operator"). We reuse
the table of names for slot lookup because it would be a waste of code
space to store the pretty name for each operator.
Previous patch c38dc3ccc7 allowed any
object to be compared with any other, using pointer comparison for a
fallback. As such, existing code which checked for this case is no
longer needed.
This is a simple optimisation inspired by JITing technology: we cache in
the bytecode (using 1 byte) the offset of the last successful lookup in
a map. This allows us next time round to check in that location in the
hash table (mp_map_t) for the desired entry, and if it's there use that
entry straight away. Otherwise fallback to a normal map lookup.
Works for LOAD_NAME, LOAD_GLOBAL, LOAD_ATTR and STORE_ATTR opcodes.
On a few tests it gives >90% cache hit and greatly improves speed of
code.
Disabled by default. Enabled for unix and stmhal ports.
Going from MICROPY_ERROR_REPORTING_NORMAL to
MICROPY_ERROR_REPORTING_TERSE now saves 2020 bytes ROM for ARM Thumb2,
and 2200 bytes ROM for 32-bit x86.
This is about a 2.5% code size reduction for bare-arm.
Uninitialised struct members get a default value of 0/false, so this is
not strictly needed. But it actually decreases code size because when
all members are initialised the compiler doesn't need to insert a call
to memset to clear everything. In other words, setting 1 extra member
to 0 uses less code than calling memset.
ROM savings in bytes: 32-bit unix: 100; bare-arm: 44; stmhal: 52.
Because (for Thumb) a function pointer has the LSB set, pointers to
dynamic functions in RAM (eg native, viper or asm functions) were not
being traced by the GC. This patch is a comprehensive fix for this.
Addresses issue #820.
Damien George