Each thread specializes a thread-local copy of the bytecode, created on the first RESUME, in free-threaded builds. All copies of the bytecode for a code object are stored in the co_tlbc array on the code object. Threads reserve a globally unique index identifying its copy of the bytecode in all co_tlbc arrays at thread creation and release the index at thread destruction. The first entry in every co_tlbc array always points to the "main" copy of the bytecode that is stored at the end of the code object. This ensures that no bytecode is copied for programs that do not use threads.
Thread-local bytecode can be disabled at runtime by providing either -X tlbc=0 or PYTHON_TLBC=0. Disabling thread-local bytecode also disables specialization.
Concurrent modifications to the bytecode made by the specializing interpreter and instrumentation use atomics, with specialization taking care not to overwrite an instruction that was instrumented concurrently.
* Spill the evaluation around escaping calls in the generated interpreter and JIT.
* The code generator tracks live, cached values so they can be saved to memory when needed.
* Spills the stack pointer around escaping calls, so that the exact stack is visible to the cycle GC.
The adaptive counter doesn't do anything currently in the free-threaded
build and TSan reports a data race due to concurrent modifications to
the counter.
This PR sets up tagged pointers for CPython.
The general idea is to create a separate struct _PyStackRef for everything on the evaluation stack to store the bits. This forces the C compiler to warn us if we try to cast things or pull things out of the struct directly.
Only for free threading: We tag the low bit if something is deferred - that means we skip incref and decref operations on it. This behavior may change in the future if Mark's plans to defer all objects in the interpreter loop pans out.
This implies a strict stack reference discipline is required. ALL incref and decref operations on stackrefs must use the stackref variants. It is unsafe to untag something then do normal incref/decref ops on it.
The new incref and decref variants are called dup and close. They mimic a "handle" API operating on these stackrefs.
Please read Include/internal/pycore_stackref.h for more information!
---------
Co-authored-by: Mark Shannon <9448417+markshannon@users.noreply.github.com>
Introduce a unified 16-bit backoff counter type (``_Py_BackoffCounter``),
shared between the Tier 1 adaptive specializer and the Tier 2 optimizer. The
API used for adaptive specialization counters is changed but the behavior is
(supposed to be) identical.
The behavior of the Tier 2 counters is changed:
- There are no longer dynamic thresholds (we never varied these).
- All counters now use the same exponential backoff.
- The counter for ``JUMP_BACKWARD`` starts counting down from 16.
- The ``temperature`` in side exits starts counting down from 64.
This change adds an `eval_breaker` field to `PyThreadState`. The primary
motivation is for performance in free-threaded builds: with thread-local eval
breakers, we can stop a specific thread (e.g., for an async exception) without
interrupting other threads.
The source of truth for the global instrumentation version is stored in the
`instrumentation_version` field in PyInterpreterState. Threads usually read the
version from their local `eval_breaker`, where it continues to be colocated
with the eval breaker bits.
This adds a safe memory reclamation scheme based on FreeBSD's "GUS" and
quiescent state based reclamation (QSBR). The API provides a mechanism
for callers to detect when it is safe to free memory that may be
concurrently accessed by readers.
This fixes a recently introduced bug where the deferred count is being unnecessarily decremented to counteract an increment elsewhere that is no longer happening. This caused the values to flip around to "very large" 64-bit numbers.
This makes the Tier 2 interpreter a little faster.
I calculated by about 3%,
though I hesitate to claim an exact number.
This starts by doubling the trace size limit (to 512),
making it more likely that loops fit in a trace.
The rest of the approach is to only load
`oparg` and `operand` in cases that use them.
The code generator know when these are used.
For `oparg`, it will conditionally emit
```
oparg = CURRENT_OPARG();
```
at the top of the case block.
(The `oparg` variable may be referenced multiple times
by the instructions code block, so it must be in a variable.)
For `operand`, it will use `CURRENT_OPERAND()` directly
instead of referencing the `operand` variable,
which no longer exists.
(There is only one place where this will be used.)
- There is no longer a separate Python/executor.c file.
- Conventions in Python/bytecodes.c are slightly different -- don't use `goto error`,
you must use `GOTO_ERROR(error)` (same for others like `unused_local_error`).
- The `TIER_ONE` and `TIER_TWO` symbols are only valid in the generated (.c.h) files.
- In Lib/test/support/__init__.py, `Py_C_RECURSION_LIMIT` is imported from `_testcapi`.
- On Windows, in debug mode, stack allocation grows from 8MiB to 12MiB.
- **Beware!** This changes the env vars to enable uops and their debugging
to `PYTHON_UOPS` and `PYTHON_LLTRACE`.
Statistics gathering is now off by default. Use the "-X pystats"
command line option or set the new PYTHONSTATS environment variable
to 1 to turn statistics gathering on at Python startup.
Statistics are no longer dumped at exit if statistics gathering was
off or statistics have been cleared.
Changes:
* Add PYTHONSTATS environment variable.
* sys._stats_dump() now returns False if statistics are not dumped
because they are all equal to zero.
* Add PyConfig._pystats member.
* Add tests on sys functions and on setting PyConfig._pystats to 1.
* Add Include/cpython/pystats.h and Include/internal/pycore_pystats.h
header files.
* Rename '_py_stats' variable to '_Py_stats'.
* Exclude Include/cpython/pystats.h from the Py_LIMITED_API.
* Move pystats.h include from object.h to Python.h.
* Add _Py_StatsOn() and _Py_StatsOff() functions. Remove
'_py_stats_struct' variable from the API: make it static in
specialize.c.
* Document API in Include/pystats.h and Include/cpython/pystats.h.
* Complete pystats documentation in Doc/using/configure.rst.
* Don't write "all zeros" stats: if _stats_off() and _stats_clear()
or _stats_dump() were called.
* _PyEval_Fini() now always call _Py_PrintSpecializationStats() which
does nothing if stats are all zeros.
Co-authored-by: Michael Droettboom <mdboom@gmail.com>
This finishes the work begun in gh-107760. When, while projecting a superblock, we encounter a call to a short, simple function, the superblock will now enter the function using `_PUSH_FRAME`, continue through it, and leave it using `_POP_FRAME`, and then continue through the original code. Multiple frame pushes and pops are even possible. It is also possible to stop appending to the superblock in the middle of a called function, when running out of space or encountering an unsupported bytecode.
* Split `CALL_PY_EXACT_ARGS` into uops
This is only the first step for doing `CALL` in Tier 2.
The next step involves tracing into the called code object and back.
After that we'll have to do the remaining `CALL` specialization.
Finally we'll have to deal with `KW_NAMES`.
Note: this moves setting `frame->return_offset` directly in front of
`DISPATCH_INLINED()`, to make it easier to move it into `_PUSH_FRAME`.
By turning `assert(kwnames == NULL)` into a macro that is not in the "forbidden" list, many instructions that formerly were skipped because they contained such an assert (but no other mention of `kwnames`) are now supported in Tier 2. This covers 10 instructions in total (all specializations of `CALL` that invoke some C code):
- `CALL_NO_KW_TYPE_1`
- `CALL_NO_KW_STR_1`
- `CALL_NO_KW_TUPLE_1`
- `CALL_NO_KW_BUILTIN_O`
- `CALL_NO_KW_BUILTIN_FAST`
- `CALL_NO_KW_LEN`
- `CALL_NO_KW_ISINSTANCE`
- `CALL_NO_KW_METHOD_DESCRIPTOR_O`
- `CALL_NO_KW_METHOD_DESCRIPTOR_NOARGS`
- `CALL_NO_KW_METHOD_DESCRIPTOR_FAST`
This effectively reverts bb578a0, restoring the original DEOPT_IF() macro in ceval_macros.h, and redefining it in the Tier 2 interpreter. We can get rid of the PREDICTED() macros there as well!
Added a new, experimental, tracing optimizer and interpreter (a.k.a. "tier 2"). This currently pessimizes, so don't use yet -- this is infrastructure so we can experiment with optimizing passes. To enable it, pass ``-Xuops`` or set ``PYTHONUOPS=1``. To get debug output, set ``PYTHONUOPSDEBUG=N`` where ``N`` is a debug level (0-4, where 0 is no debug output and 4 is excessively verbose).
All of this code is likely to change dramatically before the 3.13 feature freeze. But this is a first step.
mpage
Mark Shannon
Sam Gross
Brandt Bucher
Ken Jin
Tian Gao
Dino Viehland
Guido van Rossum
Michael Droettboom
Brett Simmers
Kirill Podoprigora
Irit Katriel
Victor Stinner