This patch adds full support for unwinding jumps to the native emitter.
This means that return/break/continue can be used in try-except,
try-finally and with statements. For code that doesn't use unwinding jumps
there is almost no overhead added to the generated code.
The native emitter keeps the current exception in a slot in its C stack
(instead of on its Python value stack), so when it catches an exception it
must explicitly clear that slot so the same exception is not reraised later
on.
Prior to this patch, native code would use a full nlr_buf_t for each
exception handler (try-except, try-finally, with). For nested exception
handlers this would use a lot of C stack and be rather inefficient.
This patch changes how exceptions are handled in native code by setting up
only a single nlr_buf_t context for the entire function, and then manages a
state machine (using the PC) to work out which exception handler to run
when an exception is raised by an nlr_jump. This keeps the C stack usage
at a constant level regardless of the depth of Python exception blocks.
The patch also fixes an existing bug when local variables are written to
within an exception handler, then their value was incorrectly restored if
an exception was raised (since the nlr_jump would restore register values,
back to the point of the nlr_push).
And it also gets nested try-finally+with working with the viper emitter.
Broadly speaking, efficiency of executing native code that doesn't use
any exception blocks is unchanged, and emitted code size is only slightly
increased for such function. C stack usage of all native functions is
either equal or less than before. Emitted code size for native functions
that use exception blocks is increased by roughly 10% (due in part to
fixing of above-mentioned bugs).
But, most importantly, this patch allows to implement more Python features
in native code, like unwind jumps and yielding from within nested exception
blocks.
There is no need to have three copies of the exception object on the top of
the native value stack. Instead, the values on the stack should be the
first two items in an nlr_buf_t: the prev pointer and the ret_val pointer.
This is all that is needed and is what the rest of the native emitter
expects is on the stack.
This patch is essentially an optimisation. Behaviour is unchanged,
although the stack layout for native exception handling now makes more
sense.
A native function allocates space on its C stack for mp_code_state_t,
followed by its Python stack, then its locals. This patch makes sure that
the native function actually starts at the start of its Python stack,
rather than at the start of mp_code_state_t (which didn't lead to any
issues so far because the mp_code_state_t is unused after the native
function sets itself up).
On x86 archs (both 32 and 64 bit) a bool return value only sets the 8-bit
al register, and the higher bits of the ax register have an undefined
value. When testing the return value of such cases it is required to just
test al for zero/non-zero. On the other hand, checking for truth or
zero/non-zero on an integer return value requires checking all bits of the
register. These two cases must be distinguished and handled correctly in
generated native code. This patch makes sure of this.
For other supported native archs (ARM, Thumb2, Xtensa) there is no such
distinction and this patch does not change anything for them.
In non-debug mode MP_OBJ_STOP_ITERATION is zero and comparing something to
zero can be done more efficiently in assembler than comparing to a non-zero
value.
Instead of emitnative.c having configuration code for each supported
architecture, and then compiling this file multiple times with different
macros defined, this patch adds a file per architecture with the necessary
code to configure the native emitter. These files then #include the
emitnative.c file.
This simplifies emitnative.c (which is already very large), and simplifies
the build system because emitnative.c no longer needs special handling for
compilation and qstr extraction.
All the asm macro names that convert a particular architecture to a generic
interface now follow the convention whereby the "destination" (usually a
register) is specified first.
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.
- Changed: ValueError, TypeError, NotImplementedError
- OSError invocations unchanged, because the corresponding utility
function takes ints, not strings like the long form invocation.
- OverflowError, IndexError and RuntimeError etc. not changed for now
until we decide whether to add new utility functions.
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
It improves readability of code and reduces the chance to make a mistake.
This patch also fixes a bug with nan-boxing builds by rounding up the
calculation of the new NSLOTS variable, giving the correct number of slots
(being 4) even if mp_obj_t is larger than the native machine size.
Instead of caching data that is constant (code_info, const_table and
n_state), store just a pointer to the underlying function object from which
this data can be derived.
This helps reduce stack usage for the case when the mp_code_state_t
structure is stored on the stack, as well as heap usage when it's stored
on the heap.
The downside is that the VM becomes a little more complex because it now
needs to derive the data from the underlying function object. But this
doesn't impact the performance by much (if at all) because most of the
decoding of data is done outside the main opcode loop. Measurements using
pystone show that little to no performance is lost.
This patch also fixes a nasty bug whereby the bytecode can be reclaimed by
the GC during execution. With this patch there is always a pointer to the
function object held by the VM during execution, since it's stored in the
mp_code_state_t structure.
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.
The 3 kinds of comprehensions are similar enough that merging their emit
functions reduces code size. Decreases in code size in bytes are:
bare-arm:24, minimal:96, unix(NDEBUG,x86-64):328, stmhal:80, esp8266:76.
