Merge branch 'main' into encodings-cache

Author: StanFromIreland

Doc/library/asyncio-task.rst

@@ -557,7 +557,7 @@ Running Tasks Concurrently
       provides stronger safety guarantees than *gather* for scheduling a nesting of subtasks:
       if a task (or a subtask, a task scheduled by a task)
       raises an exception, *TaskGroup* will, while *gather* will not,
-      cancel the remaining scheduled tasks).
+      cancel the remaining scheduled tasks.
 
    .. _asyncio_example_gather:
 

Doc/tutorial/datastructures.rst

@@ -493,6 +493,9 @@ Curly braces or the :func:`set` function can be used to create sets.  Note: to
 create an empty set you have to use ``set()``, not ``{}``; the latter creates an
 empty dictionary, a data structure that we discuss in the next section.
 
+Because sets are unordered, iterating over them or printing them can
+produce the elements in a different order than you expect.
+
 Here is a brief demonstration::
 
    >>> basket = {'apple', 'orange', 'apple', 'pear', 'orange', 'banana'}

Doc/whatsnew/3.15.rst

@@ -1286,11 +1286,11 @@ Upgraded JIT compiler
 
 Results from the `pyperformance <https://github.com/python/pyperformance>`__
 benchmark suite report
-`4-5% <https://raw.githubusercontent.com/facebookexperimental/free-threading-benchmarking/refs/heads/main/results/bm-20260110-3.15.0a3%2B-aa8578d-JIT/bm-20260110-vultr-x86_64-python-aa8578dc54df2af9daa3-3.15.0a3%2B-aa8578d-vs-base.svg>`__
+`5-6% <https://doesjitgobrrr.com/run/2026-03-11>`__
 geometric mean performance improvement for the JIT over the standard CPython
 interpreter built with all optimizations enabled on x86-64 Linux. On AArch64
 macOS, the JIT has a
-`7-8% <https://raw.githubusercontent.com/facebookexperimental/free-threading-benchmarking/refs/heads/main/results/bm-20260110-3.15.0a3%2B-aa8578d-JIT/bm-20260110-macm4pro-arm64-python-aa8578dc54df2af9daa3-3.15.0a3%2B-aa8578d-vs-base.svg>`__
+`8-9% <https://doesjitgobrrr.com/run/2026-03-11>`__
 speedup over the :ref:`tail calling interpreter <whatsnew314-tail-call-interpreter>`
 with all optimizations enabled. The speedups for JIT
 builds versus no JIT builds range from roughly 15% slowdown to over

Include/internal/pycore_backoff.h

@@ -135,6 +135,20 @@ initial_jump_backoff_counter(_PyOptimizationConfig *opt_config)
         opt_config->jump_backward_initial_backoff);
 }
 
+// This needs to be around 2-4x of JUMP_BACKWARD_INITIAL_VALUE
+// The reasoning is that we always want loop traces to form and inline
+// functions before functions themselves warm up and link to them instead
+// of inlining.
+#define RESUME_INITIAL_VALUE 8190
+#define RESUME_INITIAL_BACKOFF 6
+static inline _Py_BackoffCounter
+initial_resume_backoff_counter(_PyOptimizationConfig *opt_config)
+{
+    return make_backoff_counter(
+        opt_config->resume_initial_value,
+        opt_config->resume_initial_backoff);
+}
+
 /* Initial exit temperature.
  * Must be larger than ADAPTIVE_COOLDOWN_VALUE,
  * otherwise when a side exit warms up we may construct

Include/internal/pycore_code.h

@@ -323,6 +323,7 @@ PyAPI_FUNC(void) _Py_Specialize_ToBool(_PyStackRef value, _Py_CODEUNIT *instr);
 PyAPI_FUNC(void) _Py_Specialize_ContainsOp(_PyStackRef value, _Py_CODEUNIT *instr);
 PyAPI_FUNC(void) _Py_GatherStats_GetIter(_PyStackRef iterable);
 PyAPI_FUNC(void) _Py_Specialize_CallFunctionEx(_PyStackRef func_st, _Py_CODEUNIT *instr);
+PyAPI_FUNC(void) _Py_Specialize_Resume(_Py_CODEUNIT *instr, PyThreadState *tstate, _PyInterpreterFrame *frame);
 
 // Utility functions for reading/writing 32/64-bit values in the inline caches.
 // Great care should be taken to ensure that these functions remain correct and

Include/internal/pycore_interp_structs.h

@@ -14,6 +14,7 @@ extern "C" {
 #include "pycore_structs.h"       // PyHamtObject
 #include "pycore_tstate.h"        // _PyThreadStateImpl
 #include "pycore_typedefs.h"      // _PyRuntimeState
+#include "pycore_uop.h"           // _PyBloomFilter
 
 #define CODE_MAX_WATCHERS 8
 #define CONTEXT_MAX_WATCHERS 8
@@ -413,6 +414,9 @@ typedef struct _PyOptimizationConfig {
     uint16_t jump_backward_initial_value;
     uint16_t jump_backward_initial_backoff;
 
+    uint16_t resume_initial_value;
+    uint16_t resume_initial_backoff;
+
     // JIT optimization thresholds
     uint16_t side_exit_initial_value;
     uint16_t side_exit_initial_backoff;
@@ -972,7 +976,10 @@ struct _is {
 
     // Optimization configuration (thresholds and flags for JIT and interpreter)
     _PyOptimizationConfig opt_config;
-    struct _PyExecutorObject *executor_list_head;
+    _PyBloomFilter *executor_blooms;             // Contiguous bloom filter array
+    struct _PyExecutorObject **executor_ptrs;    // Corresponding executor pointer array
+    size_t executor_count;                       // Number of valid executors
+    size_t executor_capacity;                    // Array capacity
     struct _PyExecutorObject *executor_deletion_list_head;
     struct _PyExecutorObject *cold_executor;
     struct _PyExecutorObject *cold_dynamic_executor;

Include/internal/pycore_magic_number.h

@@ -292,6 +292,7 @@ Known values:
     Python 3.15a4 3659 (Add CALL_FUNCTION_EX specialization)
     Python 3.15a4 3660 (Change generator preamble code)
     Python 3.15a4 3661 (Lazy imports IMPORT_NAME opcode changes)
+    Python 3.15a6 3662 (Add counter to RESUME)
 
 
     Python 3.16 will start with 3700
@@ -305,7 +306,7 @@ PC/launcher.c must also be updated.
 
 */
 
-#define PYC_MAGIC_NUMBER 3661
+#define PYC_MAGIC_NUMBER 3662
 /* This is equivalent to converting PYC_MAGIC_NUMBER to 2 bytes
    (little-endian) and then appending b'\r\n'. */
 #define PYC_MAGIC_NUMBER_TOKEN \

Include/internal/pycore_opcode_metadata.h

@@ -128,8 +128,8 @@ typedef struct {
     bool cold;
     uint8_t pending_deletion;
     int32_t index;           // Index of ENTER_EXECUTOR (if code isn't NULL, below).
-    _PyBloomFilter bloom;
-    _PyExecutorLinkListNode links;
+    int32_t bloom_array_idx;        // Index in interp->executor_blooms/executor_ptrs.
+    _PyExecutorLinkListNode links;  // Used by deletion list.
     PyCodeObject *code;  // Weak (NULL if no corresponding ENTER_EXECUTOR).
 } _PyVMData;
 
@@ -157,7 +157,7 @@ typedef struct _PyExecutorObject {
 // Export for '_opcode' shared extension (JIT compiler).
 PyAPI_FUNC(_PyExecutorObject*) _Py_GetExecutor(PyCodeObject *code, int offset);
 
-void _Py_ExecutorInit(_PyExecutorObject *, const _PyBloomFilter *);
+int _Py_ExecutorInit(_PyExecutorObject *, const _PyBloomFilter *);
 void _Py_ExecutorDetach(_PyExecutorObject *);
 void _Py_BloomFilter_Init(_PyBloomFilter *);
 void _Py_BloomFilter_Add(_PyBloomFilter *bloom, void *obj);
@@ -361,6 +361,8 @@ _PyJit_TryInitializeTracing(PyThreadState *tstate, _PyInterpreterFrame *frame,
     int oparg, _PyExecutorObject *current_executor);
 
 PyAPI_FUNC(void) _PyJit_FinalizeTracing(PyThreadState *tstate, int err);
+PyAPI_FUNC(bool) _PyJit_EnterExecutorShouldStopTracing(int og_opcode);
+
 void _PyPrintExecutor(_PyExecutorObject *executor, const _PyUOpInstruction *marker);
 void _PyJit_TracerFree(_PyThreadStateImpl *_tstate);
 

Include/internal/pycore_optimizer.h

@@ -36,7 +36,7 @@ typedef struct _PyUOpInstruction{
 } _PyUOpInstruction;
 
 // This is the length of the trace we translate initially.
-#ifdef Py_DEBUG
+#if defined(Py_DEBUG) && defined(_Py_JIT)
     // With asserts, the stencils are a lot larger
 #define UOP_MAX_TRACE_LENGTH 1000
 #else