In applications that use little memory and run GC regularly, the cost of
the sweep phase quickly becomes prohibitives as the amount of RAM
increases.
On an ESP32-S3 with 2 MB of external SPIRAM, for example, a trivial GC
cycle takes a minimum of 40ms, virtually all of it in the sweep phase.
Similarly, on the UNIX port with 1 GB of heap, a trivial GC takes 47 ms,
again virtually all of it in the sweep phase.
This commit speeds up the sweep phase in the case most of the heap is empty
by keeping track of the ID of the highest block we allocated in an area
since the last GC.
The performance benchmark run on PYBV10 shows between +0 and +2%
improvement across the existing performance tests. These tests don't
really stress the GC, so they were also run with gc.threshold(30000) and
gc.threshold(10000). For the 30000 case, performance improved by up to
+10% with this commit. For the 10000 case, performance improved by at
least +10% on 6 tests, and up to +25%.
Signed-off-by: Damien George <damien@micropython.org>
Previously this was explicitly enabled on esp32/stm32/renesas/mimxrt/samd,
but didn't get a default feature level because it wasn't in py/mpconfig.h.
With this commit it's now enabled at the "extra features" level, which adds
rp2, unix-standard, windows, esp8266, webassembly, and some nrf boards.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
When MICROPY_SCHEDULER_STATIC_NODES is enabled, the logic is unchanged.
When MICROPY_SCHEDULER_STATIC_NODES is disable, sched_state is now always
initialised to MP_SCHED_IDLE when calling mp_init(). For example, the use
of mp_sched_vm_abort(), if it aborts a running scheduled function, can lead
to the scheduler starting off in a locked state when the runtime is
restarted, and then it stays locked. This commit fixes that case by
resetting sched_state.
Signed-off-by: Damien George <damien@micropython.org>
This provides similar functionality to the former zlib.DecompIO and
especially CPython's gzip.GzipFile for both compression and decompression.
This class can be used directly, and also can be used from Python to
implement (via io.BytesIO) zlib.decompress and zlib.compress, as well as
gzip.GzipFile.
Enable/disable this on all ports/boards that zlib was previously configured
for.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
There are enough places that implement __exit__ by forwarding directly to
mp_stream_close that this saves code size.
For the cases where __exit__ is a no-op, additionally make their
MP_STREAM_CLOSE ioctl handled as a no-op.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This will be replaced with a new deflate module providing the same
functionality, with an optional frozen Python wrapper providing a
replacement zlib module.
binascii.crc32 is temporarily disabled.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Follow-up to 24c02c4eb5 for when
MICROPY_ENABLE_EXTERNAL_IMPORT=0. It now needs to try both extensible and
non-extensible modules.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Prior to this fix, async for assumed the iterator expression was a simple
identifier, and used that identifier as a local to store the intermediate
iterator object. This is incorrect behaviour.
This commit fixes the issue by keeping the iterator object on the stack as
an anonymous local variable.
Fixes issue #11511.
Signed-off-by: Damien George <damien@micropython.org>
The asyncio module now has much better CPython compatibility and
deserves to be just called "asyncio".
This will avoid people having to write `from uasyncio import asyncio`.
Renames all files, and updates port manifests to use the new path. Also
renames the built-in _uasyncio to _asyncio.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
PEP-498 allows for conversion specifiers like !r and !s to convert the
expression declared in braces to be passed through repr() and str()
respectively.
This updates the logic that detects the end of the expression to also stop
when it sees "![rs]" that is either at the end of the f-string or before
the ":" indicating the start of the format specifier. The "![rs]" is now
retained in the format string, whereas previously it stayed on the end
of the expression leading to a syntax error.
Previously: `f"{x!y:z}"` --> `"{:z}".format(x!y)`
Now: `f"{x!y:z}"` --> `"{!y:z}".format(x)`
Note that "!a" is not supported by `str.format` as MicroPython has no
`ascii()`, but now this will raise the correct error.
Updated cpydiff and added tests.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
So that the delegation functions don't need to be put somewhere global,
like in mpconfigport.h. That would otherwise make it hard for extension
modules to use delegation.
Signed-off-by: Damien George <damien@micropython.org>
Starting with 2757acf6, the `top` variable in `nlr_jump()` in
`nlraarch64.c` was assigned to register `x19` by the compiler. However,
the assembly code writes over that register with
ldp x19, x20, [%0, #32]
since `%0` is now `x19`. This causes the next line
ldp lr, x9, [%0, #16]
to load the wrong values.
To fix the issue, we move the value of the `top` variable from an unknown
register to a known register at the beginning of the asm code then only use
known/hard-coded registers after that.
Fixes issue #11754.
Signed-off-by: David Lechner <david@pybricks.com>
Following how mkrules.cmake works. This makes it easy for a port to enable
frozen code, by defining FROZEN_MANIFEST in its Makefile.
Signed-off-by: Damien George <damien@micropython.org>
This is a MicroPython-specific module that existed to support the old
version of uasyncio. It's undocumented and not enabled on all ports and
takes up code size unnecessarily.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Previously sys.path could be modified by append/pop or slice assignment.
This allows `sys.path = [...]`, which can be simpler in many cases, but
also improves CPython compatibility.
It also allows sys.path to be set to a tuple which means that you can
clear sys.path (e.g. temporarily) with no allocations.
This also makes sys.path (and sys.argv for consistency) able to be disabled
via mpconfig. The unix port (and upytesthelper) require them, so they
explicitly verify that they're enabled.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Otherwise you can get into the confusing state where e.g. sys.ps1 is
enabled in config (via `MICROPY_PY_SYS_PS1_PS2`) but still doesn't actually
get enabled.
Also verify that the required delegation options are enabled in modsys.c.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
When compiling mpy-cross, there is no `sys` module, and so there will
be no entries in the `mp_builtin_module_delegation_table`.
MSVC doesn't like this, so instead pretend as if the feature isn't
enabled at all.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This replaces the previous QSTR_null entry in the globals dict which could
leak out to Python (e.g. via iteration of mod.__dict__) and could lead to
crashes.
It results in smaller code size at the expense of turning a lookup into a
loop, but the list it is looping over likely only contains one or two
elements.
To allow a module to register its custom attr function it can use the new
`MP_REGISTER_MODULE_DELEGATION` macro.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Updates any includes, and references from Makefiles/CMake.
This essentially reverts what was done long ago in commit
136b5cbd76
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
In order to keep "import umodule" working, the existing mechanism is
replaced with a simple fallback to drop the "u".
This makes importing of built-ins no longer touch the filesystem, which
makes a typical built-in import take ~0.15ms rather than 3-5ms.
(Weak links were added in c14a81662c)
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This renames the builtin-modules, such that help('modules') and printing
the module object will show "module" rather than "umodule".
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
Prior to this commit, importing a module that exists but has a syntax error
or some other problem that happens at import time would result in a
potentially-incomplete module object getting added to sys.modules.
Subsequent imports would use that object, resulting in confusing error
messages that hide the root cause of the problem.
This commit fixes that issue by removing the failed module from sys.modules
using the new NLR callback mechanism.
Note that it is still important to add the module to sys.modules while the
import is happening so that we can support circular imports just like
CPython does.
Fixes issue #967.
Signed-off-by: David Grayson <davidegrayson@gmail.com>
The changed functions now use less stack, and don't have any issues with
local variables needing to be declared volatile.
Testing on a PYBv1.0, imports (of .py, .mpy and frozen code) now use 64
less bytes of C stack per import depth.
Signed-off-by: Damien George <damien@micropython.org>
NLR buffers are usually quite large (use lots of C stack) and expensive to
push and pop. Some of the time they are only needed to perform clean up if
an exception happens, and then they re-raise the exception.
This commit allows optimizing that scenario by introducing a linked-list of
NLR callbacks that are called automatically when an exception is raised.
They are essentially a light-weight NLR handler that can implement a
"finally" block, i.e. clean-up when an exception is raised, or (by passing
`true` to nlr_pop_jump_callback) when execution leaves the scope.
Signed-off-by: Damien George <damien@micropython.org>
This can lead to duplicate initialisations if a module can be imported
via multiple names, so the module must track this itself anyway.
This reduces code size (diff is -40 bytes), and avoids special treatment of
builtin-modules-with-init with respect to sys.modules. No other builtin
modules get put into sys.modules.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
To use this:
- Create a built-in module, and add the module object as a member of the
parent module's globals dict.
- The submodule can set its `__name__` to either `QSTR_foo_dot_bar` or
`QSTR_bar`. The former requires using qstrdefs(port).h to make the qstr.
Because `bar` is a member of `foo`'s globals, it is possible to write
`import foo` and then immediately use `foo.bar` without importing it
explicitly. This means that if `bar` has an `__init__`, it will not be
called in this situation, and for that reason, sub-modules should not have
`__init__` methods. If this is required, then all initalisation for
sub-modules should be done by the top-level module's (i.e. `foo`'s)
`__init__` method.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This makes it so that sub-packages are resolved relative to their parent's
`__path__`, rather than re-resolving each parent's filesystem path.
The previous behavior was that `import foo.bar` would first re-search
`sys.path` for `foo`, then use the resulting path to find `bar`.
For already-loaded and u-prefixed modules, because we no longer need to
build the path from level to level, we no longer unnecessarily search
the filesystem. This should improve startup time.
Explicitly makes the resolving process clear:
- Loaded modules are returned immediately without touching the filesystem.
- Exact-match of builtins are also returned immediately.
- Then the filesystem search happens.
- If that fails, then the weak-link handling is applied.
This maintains the existing behavior: if a user writes `import time` they
will get time.py if it exits, otherwise the built-in utime. Whereas `import
utime` will always return the built-in.
This also fixes a regression from a7fa18c203
where we search the filesystem for built-ins. It is now only possible to
override u-prefixed builtins. This will remove a lot of filesystem stats
at startup, as micropython-specific modules (e.g. `pyb`) will no longer
attempt to look at the filesystem.
Added several improvements to the comments and some minor renaming and
refactoring to make it clearer how the import mechanism works. Overall
code size diff is +56 bytes on STM32.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
If sys.path is enabled, but empty, this will now no longer search the
filesystem. Previously an empty sys.path was equivalent to having
`sys.path=[""]`. This is a breaking change, but this behavior now matches
CPython.
This also provides an alternative mechanism to the u-prefix to force an
import of a builtin module:
```
import sys
_path = sys.path[:]
sys.path.clear()
import foo # Forces the built-in foo.
sys.path.extend(_path)
del _path
```
Code size diff is -32 bytes on PYBV11.
This work was funded through GitHub Sponsors.
Signed-off-by: Jim Mussared <jim.mussared@gmail.com>
This generalises and simplifies the code and follows CPython behaviour.
See similar change for floats in a07fc5b640.
Signed-off-by: Damien George <damien@micropython.org>
This is possible now that MP_UNARY_OP_INT_MAYBE exists.
As a consequence mp_obj_get_int now also supports user types, which was
previously possible with MP_UNARY_OP_INT but no tests existed for it.
Signed-off-by: Damien George <damien@micropython.org>
To be consistent with MP_UNARY_OP_INT_FLOAT and MP_UNARY_OP_INT_COMPLEX,
and allow int() to first check if a type supports __int__ before trying
other things (as per CPython).
Signed-off-by: Damien George <damien@micropython.org>
The code that handles inplace-operator to normal-binary-operator fallback
is moved in this commit from py/objtype.c to py/runtime.c, making it apply
to all types, not just user classes.
Signed-off-by: Damien George <damien@micropython.org>
So that user types can implement reverse operators and have them work with
str on the left-hand-side, eg `"a" + UserType()`.
Signed-off-by: Damien George <damien@micropython.org>
This adds a unary_op implementation for the dict_view type that makes
the implementation of `hash()` for these types compatible with CPython.
Signed-off-by: David Lechner <david@pybricks.com>
As per https://bugs.python.org/issue408326, the slice object should not be
hashable. Since MicroPython has an implicit fallback when the unary_op
slot is empty, we need to fill this slot.
Signed-off-by: David Lechner <david@pybricks.com>
Since converting to variable sized slots in mp_obj_type_t, we can now
reduce the code size a bit by removing mp_generic_unary_op() and the
corresponding slots where it is used. Instead we just implement the
generic `__hash__` operation in the runtime.
Signed-off-by: David Lechner <david@pybricks.com>
Changes in this commit:
- Add MICROPY_GC_HOOK_LOOP to gc_info() and gc_alloc(). Both of these can
be long running (many milliseconds) which is too long to be blocking in
some applications.
- Pass loop variable to MICROPY_GC_HOOK_LOOP(i) macro so that implementers
can use it, e.g. to improve performance by only calling a function every
X number of iterations.
- Drop outer call to MICROPY_GC_HOOK_LOOP in gc_mark_subtree().
