* Can now test for basic blocks.
* Optimize inverted comparisions.
* Optimize unary_not followed by a conditional jump.
* Added a new opcode, NOP, to keep code size constant.
* Applied NOP to previous transformations where appropriate.
Note, the NOP would not be necessary if other functions were
added to re-target jump addresses and update the co_lnotab mapping.
That would yield slightly faster and cleaner bytecode at the
expense of optimizer simplicity and of keeping it decoupled
from the line-numbering structure.
Added two predictions:
GET_ITER --> FOR_ITER
FOR_ITER --> STORE_FAST or UNPACK_SEQUENCE
Improves timings on pybench and timeit.py. Pystone results are neutral.
Applied to common cases:
COMPARE_OP is often followed by a JUMP_IF.
JUMP_IF is usually followed by POP_TOP.
Shows improved timings on PyStone, PyBench, and specific tests
using timeit.py:
python timeit.py -s "x=1" "if x==1: pass"
python timeit.py -s "x=1" "if x==2: pass"
python timeit.py -s "x=1" "if x: pass"
python timeit.py -s "x=100" "while x!=1: x-=1"
Potential future candidates:
GET_ITER predicts FOR_ITER
FOR_ITER predicts STORE_FAST or UNPACK_SEQUENCE
Also, applied missing goto fast_next_opcode to DUP_TOPX.
My previous patches should have used fast_next_opcode
in a few places instead of continue.
Also, applied one PyInt_AS_LONG macro in a place where
the type had already been checked.
-DCALL_PROFILE: Count the number of function calls executed.
When this symbol is defined, the ceval mainloop and helper functions
count the number of function calls made. It keeps detailed statistics
about what kind of object was called and whether the call hit any of
the special fast paths in the code.
Optimization:
When we take the fast_function() path, which seems to be taken for
most function calls, and there is minimal frame setup to do, avoid
call PyEval_EvalCodeEx(). The eval code ex function does a lot of
work to handle keywords args and star args, free variables,
generators, etc. The inlined version simply allocates the frame and
copies the arguments values into the frame.
The optimization gets a little help from compile.c which adds a
CO_NOFREE flag to code objects that don't have free variables or cell
variables. This change allows fast_function() to get into the fast
path with fewer tests.
I measure a couple of percent speedup in pystone with this change, but
there's surely more that can be done.
Make the code slightly shorter, faster, and easier to
read.
* Eliminate unused DUP_TOPX code for x==1.
compile.c always generates DUP_TOP instead.
* Since only two cases remain for DUP_TOPX, replace
the switch-case with if-elseif.
* The in-lined integer compare does a CheckExact on
both arguments. Since the second is a little more
likely to fail, test it first.
* The switch-case for IS/IS_NOT and IN/NOT_IN can
separate the regular and inverted cases with no
additional work. For all four paths, saves a test and
jump.
The two are semantically equivalent, but the first triggered a compiler
warning about an unused variable. Note, the preceding steps had already
accessed and decreffed the variable so the reference counts were fine.
parameter being either four or five. Currently, compile.c does not
generate calls with a parameter higher than three.
May have to be reverted if the second alpha or beta shakes out some
other tool generating this op code with a parameter of four or five.
Replaced groups of pushes and pops with indexed access to the stack and
a single adjustment (if needed) to the stacklevel.
Avoids scores of unnecessary increments and decrements to the stackpointer.
Removes unnecessary sequential dependencies so that the compiler has more
freedom for optimizations. Frees the processor for more parallel and
pipelined execution by using mostly read-only access and having few pointer
adjustments just prior to a read or write.
all along. Before instr_lb tended to be too high.
I don't think this actually makes any difference, given what the compiler
produces, but it makes me a bit happier.
patch #617312, both on the trunk and the 22-maint branch.
Also added a test case, and ported the test_trace I wrote for HEAD
to 2.2.2 (with all those horrible extra 'line' events ;-).
than when this interval was first established. Checking too frequently just
adds needless overhead because most of the time there is nothing to do and
no other threads ready to run.
globals, _Py_Ticker and _Py_CheckInterval. This also implements Jeremy's
shortcut in Py_AddPendingCall that zeroes out _Py_Ticker. This allows the
test in the main loop to only test a single value.
The gory details are at
http://python.org/sf/602191
Use a slightly different strategy to determine when not to call the line
trace function. This removes the need for the RETURN_NONE opcode, so
that's gone again. Update docs and comments to match.
Thanks to Neal and Armin!
Also add a test suite. This should have come with the original patch...
in LOAD_GLOBAL. Besides saving a C function call, it saves checks
whether f_globals and f_builtins are dicts, and extracting and testing
the string object's hash code is done only once. We bail out of the
inlining if the name is not exactly a string, or when its hash is -1;
because of interning, neither should ever happen. I believe interning
guarantees that the hash code is set, and I believe that the 'names'
tuple of a code object always contains interned strings, but I'm not
assuming that -- I'm simply testing hash != -1.
On my home machine, this makes a pystone variant with new-style
classes and slots run at the same speed as classic pystone! (With
new-style classes but without slots, it is still a lot slower.)
Also, don't handle METH_OLDARGS on the fast path. All the interesting
builtins have been converted to use METH_NOARGS, METH_O, or
METH_VARARGS.
Result is another 1-2% speedup. If I can cobble together 10 of these,
it might make a difference.
This makes the code much easier to ready, because it is at a sane
indentation level. On my box this shows a 1-2% speedup, which means
nothing, except that I'm not going to worry about the performance
effects of the change.
nothing special done if keyword arguments were present, so test for
that earlier and fall through to the normal case if there are any.
This ought to slow down CFunction calls with keyword args, but I don't
care; it's a tiny (1%) improvement for pystone.
[ 587993 ] SET_LINENO killer
Remove SET_LINENO. Tracing is now supported by inspecting co_lnotab.
Many sundry changes to document and adapt to this change.
The staticforward define was needed to support certain broken C
compilers (notably SCO ODT 3.0, perhaps early AIX as well) botched the
static keyword when it was used with a forward declaration of a static
initialized structure. Standard C allows the forward declaration with
static, and we've decided to stop catering to broken C compilers. (In
fact, we expect that the compilers are all fixed eight years later.)
I'm leaving staticforward and statichere defined in object.h as
static. This is only for backwards compatibility with C extensions
that might still use it.
XXX I haven't updated the documentation.
This was a simple typo. Strange that the compiler didn't catch it!
Instead of WHY_CONTINUE, two tests used CONTINUE_LOOP, which isn't a
why_code at all, but an opcode; but even though 'why' is declared as
an enum, comparing it to an int is apparently not even worth a
warning -- not in gcc, and not in VC++. :-(
Will fix in 2.2 too.
[ 558249 ] softspace vs --disable-unicode
And #endif was in the wrong place.
Bugfix candidate, almost surely.
I think I will embark on squashing test failures in --disable-unicode builds --
a Real Bug was hiding under them.
SF bug 535905 (Evil Trashcan and GC interaction).
The SETLOCAL() macro should not DECREF the local variable in-place and
then store the new value; it should copy the old value to a temporary
value, then store the new value, and then DECREF the temporary value.
This is because it is possible that during the DECREF the frame is
accessed by other code (e.g. a __del__ method or gc.collect()) and the
variable would be pointing to already-freed memory.
BUGFIX CANDIDATE!
This fixes the symptom, but PRINT_ITEM has no way to know what (if
anything) PyFile_WriteObject() writes unless the object being printed
is a string. When the object isn't a string, this fix retains the
guess that softspace should be set after PyFile_WriteObject().
We might want to say that it's the job of filelike-object write methods
to leave the file's softspace in the correct state. That would probably
be better -- but everyone relies on PRINT_ITEM to guess for them now.
eval_frame(): Under -Qnew, INPLACE_DIVIDE wasn't getting handed off to
INPLACE_TRUE_DIVIDE (like BINARY_DIVIDE was getting handed off to
BINARY_TRUE_DIVIDE).
Bugfix candidate.
Based on the patch from Danny Yoo. The fix is in exec_statement() in
ceval.c.
There are also changes to introduce use of PyCode_GetNumFree() in
several places.
Had nothing to do with rich comparisons -- some stack cleanup code was
lost as a result of merging in Neil Schemenauer's generators patch.
Reinserted the stack cleanup code, skipping it when yielding.
leak when a class defined a __metaclass__. This fixes the problem
reported on python-dev by Ping; I dunno if it's the same as SF bug
#489669 (since that mentions Unicode).
Big Hammer to implement -Qnew as PEP 238 says it should work (a global
option affecting all instances of "/").
pydebug.h, main.c, pythonrun.c: define a private _Py_QnewFlag flag, true
iff -Qnew is passed on the command line. This should go away (as the
comments say) when true division becomes The Rule. This is
deliberately not exposed to runtime inspection or modification: it's
a one-way one-shot switch to pretend you're using Python 3.
ceval.c: when _Py_QnewFlag is set, treat BINARY_DIVIDE as
BINARY_TRUE_DIVIDE.
test_{descr, generators, zipfile}.py: fiddle so these pass under
-Qnew too. This was just a matter of s!/!//! in test_generators and
test_zipfile. test_descr was trickier, as testbinop() is passed
assumptions that "/" is the same as calling a "__div__" method; put
a temporary hack there to call "__truediv__" instead when the method
name is "__div__" and 1/2 evaluates to 0.5.
Three standard tests still fail under -Qnew (on Windows; somebody
please try the Linux tests with -Qnew too! Linux runs a whole bunch
of tests Windows doesn't):
test_augassign
test_class
test_coercion
I can't stay awake longer to stare at this (be my guest). Offhand
cures weren't obvious, nor was it even obvious that cures are possible
without major hackery.
Question: when -Qnew is in effect, should calls to __div__ magically
change into calls to __truediv__? See "major hackery" at tail end of
last paragraph <wink>.
PyEval_EvalCodeEx(): increment tstate->recursion_depth around the
decref of the frame, because the C stack for this call is still in
use and the decref can lead to __del__ methods getting called.
While this gives tstate->recursion_depth a value proportional to the
depth of the C stack (instead of a small constant no matter how
deeply __del__s recurse), it's not enough to stop the reported crash
when using the default recursion limit on Windows.
Bugfix candidate.
This patch boosts performance for comparing identical string object
by some 20% on my machine while not causing any noticable slow-down
for other operations (according to tests done with pybench).
The new profiler event stream includes a "return" event even when an
exception is being propogated, but the machinery that called the profile
hook did not save & restore the exception. In debug mode, the exception
was detected during the execution of the profile callback, which did not
have the proper internal flags set for the exception. Saving & restoring
the exception state solves the problem.
The profiler does not need to know anything about the exception state,
so we no longer call it when an exception is raised. We do, however,
make sure we *always* call the profiler when we exit a frame. This
ensures that timing events are more easily isolated by a profiler and
finally clauses that do a lot of work don't have their time
mis-allocated.
When an exception is propogated out of the frame, the C callback for
the profiler now receives a PyTrace_RETURN event with an arg of NULL;
the Python-level profile hook function will see a 'return' event with
an arg of None. This means that from Python it is impossible for the
profiler to determine if the frame exited with an exception or if it
returned None, but this doesn't matter for profiling. A C-based
profiler could tell the difference, but this doesn't seem important.
ceval.c:eval_frame(): Simplify the code in two places so that the
profiler is called for every exit from a frame
and not for exceptions.
sysmodule.c:profile_trampoline(): Make sure we don't expose Python
code to NULL; use None instead.
It's possible for PyErr_NormalizeException() to set the traceback
pointer to NULL. I'm not sure how to provoke this directly from
Python, although it may be possible. The error occurs when an
exception is set using PyErr_SetObject() and another exception occurs
while PyErr_NormalizeException() is creating the exception instance.
XXX As a result of this change, it's possible for an exception to
occur but sys.last_traceback to be left undefined. Not sure if this
is a problem.
popped frame-block. What an embarrassing bug! Especially for Jeremy, since
he accepted the patch :-)
This fixes SF bugs #463359 and #462937, and possibly other, *very* obscure
bugs with very deeply nested loops that continue the loop and then break out
of it or raise an exception.
compatibility, this required all places where an array of "struct
memberlist" structures was declared that is referenced from a type's
tp_members slot to change the type of the structure to PyMemberDef;
"struct memberlist" is now only used by old code that still calls
PyMember_Get/Set. The code in PyObject_GenericGetAttr/SetAttr now
calls the new APIs PyMember_GetOne/SetOne, which take a PyMemberDef
argument.
As examples, I added actual docstrings to the attributes of a few
types: file, complex, instance method, super, and xxsubtype.spamlist.
Also converted the symtable to new style getattr.
backwards compatibility. When using the class of the first base as
the metaclass, use its __class__ attribute in preference over its
ob_type slot. This ensures that we can still use classic classes as
metaclasse, as shown in the original "Metaclasses" essay. This also
makes all the examples in Demo/metaclasses/ work again (maybe these
should be turned into a test suite?).
Change to get/set/del slice operations so that if the object doesn't
support slicing, *or* if either of the slice arguments is not an int
or long, we construct a slice object and call the get/set/del item
operation instead. This makes it possible to design classes that
support slice arguments of non-integral types.
CO_FUTURE_DIVISION flag. Redid this to use Jeremy's PyCF_MASK #define
instead, so we dont have to remember to fiddle individual feature names
here again.
pythonrun.h: Also #define a PyCF_MASK_OBSOLETE mask. This isn't used
yet, but will be as part of the PEP 264 implementation (compile() mustn't
raise an error just because old code uses a flag name that's become
obsolete; a warning may be appropriate, but not an error; so compile() has
to know about obsolete flags too, but nobody is going to remember to
update compile() with individual obsolete flag names across releases either
-- i.e., this is the flip side of PyEval_MergeCompilerFlags's oversight).
The descr changes moved the dispatch for calling objects from
call_object() in ceval.c to PyObject_Call() in abstract.c.
call_object() and the many functions it used in ceval.c were no longer
used, but were not removed.
Rename meth_call() as PyCFunction_Call() so that it can be called by
the CALL_FUNCTION opcode in ceval.c.
Also, fix error message that referred to PyEval_EvalCodeEx() by its
old name eval_code2(). (I'll probably refer to it by its old name,
too.)
Neil Schemenauer
Raymond Hettinger
Michael W. Hudson
Mark Hammond
Guido van Rossum
Just van Rossum
Jeremy Hylton
Neal Norwitz
Skip Montanaro
Fred Drake
Tim Peters
Martin v. Löwis
Marc-André Lemburg
Thomas Wouters
Sjoerd Mullender