Documentation build overview

📚 cpython-previews | 🛠️ Build #32465483 | 📁 Comparing 01fe604 against main (d4eee16)

  🔍 Preview build  

🤖 New build scheduled with the buildbot fleet by @markshannon for commit 01fe604 🤖

Results will be shown at:

https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F149103%2Fmerge

If you want to schedule another build, you need to add the 🔨 test-with-buildbots label again.

🤖 New build scheduled with the buildbot fleet by @markshannon for commit 98073f5 🤖

Results will be shown at:

https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F149103%2Fmerge

If you want to schedule another build, you need to add the 🔨 test-with-buildbots label again.

@pablogsal
Can you tell what is going on with the buildbot/AMD64 CentOS9 NoGIL PR buildbot?
The test_python_calls_appear_in_the_stack_if_perf_activated test is failing.
Interstingly py_trampoline_evaluator appears 1000s of times in the test output.
Isn't that supposed to show as the Python function in perf?

@pablogsal

Can you tell what is going on with the buildbot/AMD64 CentOS9 NoGIL PR buildbot?

The test_python_calls_appear_in_the_stack_if_perf_activated test is failing.

Interstingly py_trampoline_evaluator appears 1000s of times in the test output.

Isn't that supposed to show as the Python function in perf?

Perf should add the Python function below that one, yes. I need time to investigate I will try to do it this week but I am a bit overwhelmed with pending things. I will try to take a look as soon as possible.

@pablogsal
Can you tell what is going on with the buildbot/AMD64 CentOS9 NoGIL PR buildbot?
The test_python_calls_appear_in_the_stack_if_perf_activated test is failing.
Interstingly py_trampoline_evaluator appears 1000s of times in the test output.
Isn't that supposed to show as the Python function in perf?

I spent several hours digging into this one and was able to reproduce theCentOS9 NoGIL failure locally in a CentOS Stream 9 podman container with GCC 11.5 and perf 5.14.

The short version is: the PR itself did not break perf-map generation. The generated perf map still contains the py::foo, py::bar, and py::baz entries. What broke is perf's ability to unwind into those generated trampoline
frames on this builder.

On the failing build, perf script showed lots of py_trampoline_evaluator frames, but none of the generated py::foo/bar/baz symbols. That means the trampoline machinery was active, but perf could not walk the frame-pointer chain far enough to resolve the generated Python frames. The regression comes from the change in _Py_get_machine_stack_pointer(). Before this PR, it used:

__builtin_frame_address(0)

After this PR, on x86-64 it reads the real stack pointer with inline asm:

__asm__("{movq %%rsp, %0" : "=r" (result));

The new behavior is the right semantic direction for the stack-pointer work, but the old __builtin_frame_address(0) had an accidental side effect: it forced GCC to materialize an rbp frame pointer in _PyEval_EvalFrameDefault().

That matters because the perf trampoline test depends on frame-pointer unwinding.

On the failing CentOS9/GCC 11.5 build, _PyEval_EvalFrameDefault() also has this function-specific optimization attribute:

__attribute__((optimize ("no-tree-slp-vectorize")))

With the new %rsp helper, GCC generated _PyEval_EvalFrameDefault() without the normal frame-pointer prologue:

_PyEval_EvalFrameDefault:
    push   %r15
    push   %r14
    push   %r13
    push   %r12
    push   %rbx
    sub    $0x180,%rsp
    ...
    mov    %rsp,%rax

With the old __builtin_frame_address(0) helper, the same build generated:

_PyEval_EvalFrameDefault:
    push   %rbp
    mov    %rsp,%rbp
    push   %r15
    ...

So before this PR, the test passed because __builtin_frame_address(0) was preserving the frame-pointer chain. After this PR, that accidental protection disappeared, and perf's frame-pointer unwinder could no longer walk through _PyEval_EvalFrameDefault() into the generated trampoline frame.

The fix is to keep the new real stack-pointer behavior, but make the eval-loop optimization attribute explicitly preserve frame pointers too:

#define DONT_SLP_VECTORIZE \
    __attribute__((optimize ("no-tree-slp-vectorize", "no-omit-frame-pointer")))

I verified this in the CentOS Stream 9 reproduction:

1. New %rsp helper plus the old DONT_SLP_VECTORIZE: reproduces the buildbot failure.
2. Reverting only _Py_get_machine_stack_pointer() to__builtin_frame_address(0): test passes.
3. New %rsp helper plus optimize("no-tree-slp-vectorize", "no-omit-frame-pointer"): test passes.
4. Disassembly after the fix again shows _PyEval_EvalFrameDefault() starting with:

push   %rbp
mov    %rsp,%rbp