GH-133231: Changes to executor management to support proposed sys._jit module

[markshannon]

This PR:

• Tracks the current executor on the thread-state, not the previous one
• Sets the current executor to NULL when entering PyEval_EvalDefault and resetting it when leaving.
• Batch executors for deallocation to avoid having to constantly incref executors; this is an ad-hoc form of deferred reference counting.

Should simplify the implementation of sys._jit.is_enabled() as it is simply tstate->current_executor != NULL and doesn't require additional state to be tracked.

• Issue: Add JIT utilities in sys._jit #133231

[markshannon]

The msvc jit failure might be fixed by #132746

[Fidget-Spinner]

The msvc jit failure might be fixed by #132746

Wow I did not expect that to fix it.

[brandtbucher]

I think the overall approach is neat, but I think there are a couple of potential issues (and possible cleanups) here.

Long story short, anything that uses the entry_frame for this purpose is going to run into issues with the tail-calling interpreter (whether it's your approach, or mine). It's a hard problem.

[brandtbucher]

Is this needed? We only need one slot in localsplus, which _PyInterpreterFrame already has.

[markshannon]

We now need two slots. One on the stack for the return value, as before, and one local variable to hold the current executor.

[brandtbucher]

Ah, I see. Forgot about the return value.

Maybe add a comment that the one stack slot is for the return value, and the one local is for the executor.

[bedevere-app]

When you're done making the requested changes, leave the comment: I have made the requested changes; please review again.

[brandtbucher]

Idead: if we have a struct struct { PyExecutorObject *executor, struct node *prev} node;, we can stack allocate one of these in every GOTO_TIER_TWO call and restore it afterwards. Then keep the top on the tstate.

Something like this:

diff --git a/Python/ceval_macros.h b/Python/ceval_macros.h
index e1d2673848c..c77c2cafee7 100644
--- a/Python/ceval_macros.h
+++ b/Python/ceval_macros.h
@@ -360,10 +360,13 @@ do {                                                   \
     OPT_STAT_INC(traces_executed);                     \
     _PyExecutorObject *_executor = (EXECUTOR);         \
     jit_func jitted = _executor->jit_code;             \
+    node prev = tstate->current_executor;              \
+    tstate->current_executor = (node){_executor, &prev}; \
     /* Keep the shim frame alive via the executor: */  \
     Py_INCREF(_executor);                              \
     next_instr = jitted(frame, stack_pointer, tstate); \
     Py_DECREF(_executor);                              \
+    tstate->current_executor = prev;                   \
     Py_CLEAR(tstate->previous_executor);               \
     frame = tstate->current_frame;                     \
     stack_pointer = _PyFrame_GetStackPointer(frame);   \
@@ -379,6 +382,9 @@ do { \
     OPT_STAT_INC(traces_executed); \
     next_uop = (EXECUTOR)->trace; \
     assert(next_uop->opcode == _START_EXECUTOR); \
+    /* prev is declared local to the tier two interpreter: */ \
+    prev = tstate->current_executor; \
+    tstate->current_executor = (node){_executor, &prev}; \
     goto enter_tier_two; \
 } while (0)
 #endif
@@ -386,6 +392,7 @@ do { \
 #define GOTO_TIER_ONE(TARGET)                                         \
     do                                                                \
     {                                                                 \
+        tstate->current_executor = prev;                              \
         next_instr = (TARGET);                                        \
         OPT_HIST(trace_uop_execution_counter, trace_run_length_hist); \
         _PyFrame_SetStackPointer(frame, stack_pointer);               \

No entry frame needed.

[markshannon]

I don't see how that would work, as macros don't have scope. But its moot anyway as we can use the entry frame.

The entry frame exists in the tail-calling interpreter, it is the entry variable that we can't use. But we don't need to. Anywhere we return, either in INTERPRETER_EXIT or exit_unwind, the frame is the entry frame: frame == &entry.frame.

[markshannon]

I have made the requested changes; please review again

[bedevere-app]

Thanks for making the requested changes!

@brandtbucher: please review the changes made to this pull request.

[brandtbucher]

I don't see how that would work, as macros don't have scope.

I mean, it would just be a local declared in the _PyEval_EvalFrameDefault itself for the tail-calling interpreter, and would become local to whatever bytecode is using the macro on other builds.

But as you say...

But its moot anyway as we can use the entry frame.

The entry frame exists in the tail-calling interpreter, it is the entry variable that we can't use. But we don't need to. Anywhere we return, either in INTERPRETER_EXIT or exit_unwind, the frame is the entry frame: frame == &entry.frame.

Ah, this is what I missed. Makes sense: we're always able to set it when entering because we have entry handy. And we're always able to reset it when exiting because it's just the current frame. Thanks for explaining.

[brandtbucher]

Looks good, thanks. I'm just going to kick off some JIT benchmarks to be safe.

[brandtbucher]

Minor: this seems redundant, given that we're setting it one line above.

[brandtbucher]

You can either do this here, or in the second half of the _Py_JIT implementation of GOTO_TIER_TWO in ceval_macros.h, after the JIT code returns. No need to do it in both places.

[brandtbucher]

Side note, this instruction now exists solely to set the current_executor in the tier two interpreter. We should try to remove it as a follow-up (doesn't need to be in this PR).

[markshannon]

We can replace all uses of current_executor with an operand, which will remove the current_executor local variable from the tier 2 interpreter and the _JIT_EXECUTOR patch from the JIT template.

We can also shrink _CHECK_VALIDITY a bit by replacing current_executor->vm_data.valid with *validity_ptr:

tier2 op(_CHECK_VALIDITY, (validity_ptr/4 --)) {
    DEOPT_IF(*validity_ptr == 0);
}

For another PR.

[brandtbucher]

Looks fine.