With the previous patch combining 3 emit functions into 1, it now makes
sense to also combine the corresponding VM opcodes, which is what this
patch does. This eliminates 2 opcodes which simplifies the VM and reduces
code size, in bytes: bare-arm:44, minimal:64, unix(NDEBUG,x86-64):272,
stmhal:92, esp8266:200. Profiling (with a simple script that creates many
list/dict/set comprehensions) shows no measurable change in performance.
These can be used to insert arbitrary checks, polling, etc into the VM.
They are left general because the VM is a highly tuned loop and it should
be up to a given port how that port wants to modify the VM internals.
One common use would be to insert a polling check, but only done after
a certain number of opcodes were executed, so as not to slow down the VM
too much. For example:
#define MICROPY_VM_HOOK_COUNT (30)
#define MICROPY_VM_HOOK_INIT static uint vm_hook_divisor = MICROPY_VM_HOOK_COUNT
#define MICROPY_VM_HOOK_POLL if (--vm_hook_divisor == 0) { \
vm_hook_divisor = MICROPY_VM_HOOK_COUNT;
extern void vm_hook_function(void);
vm_hook_function();
}
#define MICROPY_VM_HOOK_LOOP MICROPY_VM_HOOK_POLL
#define MICROPY_VM_HOOK_RETURN MICROPY_VM_HOOK_POLL
Fixes#1684 and makes "not" match Python semantics. The code is also
simplified (the separate MP_BC_NOT opcode is removed) and the patch saves
68 bytes for bare-arm/ and 52 bytes for minimal/.
Previously "not x" was implemented as !mp_unary_op(x, MP_UNARY_OP_BOOL),
so any given object only needs to implement MP_UNARY_OP_BOOL (and the VM
had a special opcode to do the ! bit).
With this patch "not x" is implemented as mp_unary_op(x, MP_UNARY_OP_NOT),
but this operation is caught at the start of mp_unary_op and dispatched as
!mp_obj_is_true(x). mp_obj_is_true has special logic to test for
truthness, and is the correct way to handle the not operation.
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.
Main changes when MICROPY_PERSISTENT_CODE is enabled are:
- qstrs are encoded as 2-byte fixed width in the bytecode
- all pointers are removed from bytecode and put in const_table (this
includes const objects and raw code pointers)
Ultimately this option will enable persistence for not just bytecode but
also native code.
Previous to this patch each time a bytes object was referenced a new
instance (with the same data) was created. With this patch a single
bytes object is created in the compiler and is loaded directly at execute
time as a true constant (similar to loading bignum and float objects).
This saves on allocating RAM and means that bytes objects can now be
used when the memory manager is locked (eg in interrupts).
The MP_BC_LOAD_CONST_BYTES bytecode was removed as part of this.
Generated bytecode is slightly larger due to storing a pointer to the
bytes object instead of the qstr identifier.
Code size is reduced by about 60 bytes on Thumb2 architectures.
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__".
When generator raises exception, it is automatically terminated (by setting
its code_state.ip to 0), which interferes with this check.
Triggered in particular by CPython's test_pep380.py.
This doesn't handle case fo enclosed except blocks, but once again,
sys.exc_info() support is a workaround for software which uses it
instead of properly catching exceptions via variable in except clause.
The implementation is very basic and non-compliant and provided solely for
CPython compatibility. The function itself is bad Python2 heritage, its
usage is discouraged.
Before this patch a "with" block needed to create a bound method object
on the heap for the __exit__ call. Now it doesn't because we use
load_method instead of load_attr, and save the method+self on the stack.
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.
I.e. in this mode, C stack will never be used to call a Python function,
but if there's no free heap for a call, it will be reported as
RuntimeError (as expected), not MemoryError.
Previous to this patch, a big-int, float or imag constant was interned
(made into a qstr) and then parsed at runtime to create an object each
time it was needed. This is wasteful in RAM and not efficient. Now,
these constants are parsed straight away in the parser and turned into
objects. This allows constants with large numbers of digits (so
addresses issue #1103) and takes us a step closer to #722.
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.
This patch consolidates all global variables in py/ core into one place,
in a global structure. Root pointers are all located together to make
GC tracing easier and more efficient.
This is for efficiency, so we don't need to subtract 1 from the ip
before storing it to code_state->ip. It saves a lot of ROM bytes on
unix and stmhal.
Mirroring ip to a volatile memory variable for each opcode is an expensive
operation. For quite a lot of often executed opcodes like stack manipulation
or jumps, exceptions cannot actually happen. So, record ip only for opcode
where that's possible.
This optimisation reduces the VM exception stack element (mp_exc_stack_t)
by 1 word, by using bit 1 of a pointer to store whether the opcode was a
FINALLY or WITH opcode. This optimisation was pending, waiting for
maturity of the exception handling code, which has now proven itself.
Saves 1 machine word RAM for each exception (4->3 words per exception).
Increases stmhal code by 4 bytes, and decreases unix x64 code by 32
bytes.
This allows to implement KeyboardInterrupt on unix, and a much safer
ctrl-C in stmhal port. First ctrl-C is a soft one, with hope that VM
will notice it; second ctrl-C is a hard one that kills anything (for
both unix and stmhal).
One needs to check for a pending exception in the VM only for jump
opcodes. Others can't produce an infinite loop (infinite recursion is
caught by stack check).
There is a lot potential in compress bytecodes and make more use of the
coding space. This patch introduces "multi" bytecodes which have their
argument included in the bytecode (by addition).
UNARY_OP and BINARY_OP now no longer take a 1 byte argument for the
opcode. Rather, the opcode is included in the first byte itself.
LOAD_FAST_[0,1,2] and STORE_FAST_[0,1,2] are removed in favour of their
multi versions, which can take an argument between 0 and 15 inclusive.
The majority of LOAD_FAST/STORE_FAST codes fit in this range and so this
saves a byte for each of these.
LOAD_CONST_SMALL_INT_MULTI is used to load small ints between -16 and 47
inclusive. Such ints are quite common and now only need 1 byte to
store, and now have much faster decoding.
In all this patch saves about 2% RAM for typically bytecode (1.8% on
64-bit test, 2.5% on pyboard test). It also reduces the binary size
(because bytecodes are simplified) and doesn't harm performance.
Code-info size, block name, source name, n_state and n_exc_stack now use
variable length encoded uints. This saves 7-9 bytes per bytecode
function for most functions.
With a file with 1 line (and an error on that line), used to show the
line as number 0. Now shows it correctly as line number 1.
But, when line numbers are disabled, it now prints line number 1 for any
line that has an error (instead of 0 as previously). This might end up
being confusing, but requires extra RAM and/or hack logic to make it
print something special in the case of no line numbers.
Reduces by about a factor of 10 on average the amount of RAM needed to
store the line-number to bytecode map in the bytecode prelude.
Using CPython3.4's stdlib for statistics: previously, an average of
13 bytes were used per (bytecode offset, line-number offset) pair, and
now with this improvement, that's down to 1.3 bytes on average.
Large RAM usage before was due to some very large steps in line numbers,
both from the start of the first line in a function way down in the
file, and also functions that have big comments and/or big strings in
them (both cases were significant).
Although the savings are large on average for the CPython stdlib, it
won't have such a big effect for small scripts used in embedded
programming.
Addresses issue #648.
This may seem a bit of a risky change, in that it may introduce crazy
bugs with respect to volatile variables in the VM loop. But, I think it
should be fine: code_state points to some external memory, so the
compiler should always read/write to that memory when accessing the
ip/sp variables (ie not put them in registers).
Anyway, it passes all tests and improves on all efficiency fronts: about
2-4% faster (64-bit unix), 16 bytes less stack space per call (64-bit
unix) and slightly less executable size (unix and stmhal).
The reason it's more efficient is save_ip and save_sp were volatile
variables, so were anyway stored on the stack (in memory, not regs).
Thus converting them to code_state->{ip, sp} doesn't cost an extra
memory dereference (except maybe to get code_state, but that can be put
in a register and then made more efficient for other uses of it).
Conflicts:
py/vm.c
Fixed stack underflow check. Use UINT_FMT/INT_FMT where necessary.
Specify maximum VM-stack byte size by multiple of machine word size, so
that on 64 bit machines it has same functionality as 32 bit.
This improves stack usage in callers to mp_execute_bytecode2, and is step
forward towards unifying execution interface for function and generators
(which is important because generators don't even support full forms
of arguments passing (keywords, etc.)).
Needed to pop the iterator object when breaking out of a for loop. Need
also to be careful to unwind exception handler before popping iterator.
Addresses issue #635.
This helps the compiler do its optimisation, makes it clear which
variables are local per opcode and which global, and makes it consistent
when extra variables are needed in an opcode (in addition to old obj1,
obj2 pair, for example).
Could also make unum local, but that's for another time.
Blanket wide to all .c and .h files. Some files originating from ST are
difficult to deal with (license wise) so it was left out of those.
Also merged modpyb.h, modos.h, modstm.h and modtime.h in stmhal/.
3 emitter functions are needed only for emitcpy, and so we can #if them
out when compiling with emitcpy support.
Also remove unused SETUP_LOOP bytecode.
Closed over variables are now passed on the stack, instead of creating a
tuple and passing that. This way memory for the closed over variables
can be allocated within the closure object itself. See issue #510 for
background.
On stmhal, computed gotos make the binary about 1k bigger, but makes it
run faster, and we have the room, so why not. All tests pass on
pyboard using computed gotos.
Things get tricky when using the nlr code to catch exceptions. Need to
ensure that the variables (stack layout) in the exception handler are
the same as in the bit protected by the exception handler.
Prior to this patch there were a few bugs. 1) The constant
mp_const_MemoryError_obj was being preloaded to a specific location on
the stack at the start of the function. But this location on the stack
was being overwritten in the opcode loop (since it didn't think that
variable would ever be referenced again), and so when an exception
occurred, the variable holding the address of MemoryError was corrupt.
2) The FOR_ITER opcode detection in the exception handler used sp, which
may or may not contain the right value coming out of the main opcode
loop.
With this patch there is a clear separation of variables used in the
opcode loop and in the exception handler (should fix issue (2) above).
Furthermore, nlr_raise is no longer used in the opcode loop. Instead,
it jumps directly into the exception handler. This tells the C compiler
more about the possible code flow, and means that it should have the
same stack layout for the exception handler. This should fix issue (1)
above. Indeed, the generated (ARM) assembler has been checked explicitly,
and with 'goto exception_handler', the problem with &MemoryError is
fixed.
This may now fix problems with rge-sm, and probably many other subtle
bugs yet to show themselves. Incidentally, rge-sm now passes on
pyboard (with a reduced range of integration)!
Main lesson: nlr is tricky. Don't use nlr_push unless you know what you
are doing! Luckily, it's not used in many places. Using nlr_raise/jump
is fine.
Attempt to address issue #386. unique_code_id's have been removed and
replaced with a pointer to the "raw code" information. This pointer is
stored in the actual byte code (aligned, so the GC can trace it), so
that raw code (ie byte code, native code and inline assembler) is kept
only for as long as it is needed. In memory it's now like a tree: the
outer module's byte code points directly to its children's raw code. So
when the outer code gets freed, if there are no remaining functions that
need the raw code, then the children's code gets freed as well.
This is pretty much like CPython does it, except that CPython stores
indexes in the byte code rather than machine pointers. These indices
index the per-function constant table in order to find the relevant
code.
This is necessary to catch all cases where locals are referenced before
assignment. We still keep the _0, _1, _2 versions of LOAD_FAST to help
reduced the byte code size in RAM.
Addresses issue #457.
This simplifies the compiler a little, since now it can do 1 pass over
a function declaration, to determine default arguments. I would have
done this originally, but CPython 3.3 somehow had the default keyword
args compiled before the default position args (even though they appear
in the other order in the text of the script), and I thought it was
important to have the same order of execution when evaluating default
arguments. CPython 3.4 has changed the order to the more obvious one,
so we can also change.
Damien George
Paul Sokolovsky
stijn
Marcus von Appen
AZ Huang