GH-126910: add test for manual frame unwinding

[diegorusso]

Add a test that unwinds the stack using frame pointers. We test on any platform that frame pointers are enable (the others will be skipped). It tests with and without the JIT.

Output with PYTHON_JIT=0

#00 0xbcb53b0332f0 -> python
#01 0xbcb53b1c42b8 -> python
#02 0xbcb53b1ce4b8 -> python
#03 0xbcb53b1d758c -> python
#04 0xbcb53b0332f0 -> python
#05 0xbcb53b1c4774 -> python
#06 0xbcb53b1cfaec -> python
...
#68 0xbe2cbdec495c -> python
#69 0xeaad695b259c -> other
#70 0x60eaad695b267c -> other
#71 0x6be2cbdbc63f0 -> other
{"length": 53, "python_frames": 50, "jit_frames": 0, "other_frames": 3, "jit_code_ranges": 0, "jit_backend": "jit"}

Output with PYTHON_JIT=1

JIT ranges:
  0xeaad6985c000-0xeaad6985e000
#00 0xbe2cbdc332f0 -> python
#01 0xbe2cbddc42b8 -> python
#02 0xbe2cbddce4b8 -> python
#03 0xbe2cbddd758c -> python
#04 0xbe2cbdc332f0 -> python
#05 0xbe2cbddc4774 -> python
#06 0xeaad6985c898 -> jit
...
#61 0xeaad6985b050 -> jit
#62 0xbe2cbddd42a4 -> python
#63 0xbe2cbddd72d4 -> python
#64 0xbe2cbde90d3c -> python
#65 0xbe2cbde910d0 -> python
#66 0xbe2cbde93490 -> python
#67 0xbe2cbdec3d30 -> python
#68 0xbe2cbdec495c -> python
#69 0xeaad695b259c -> other
#70 0x60eaad695b267c -> other
#71 0x6be2cbdbc63f0 -> other
{"length": 72, "python_frames": 49, "jit_frames": 20, "other_frames": 3, "jit_code_ranges": 1, "jit_backend": "jit"}

The core of the test has been taken from @pablogsal repository https://github.com/pablogsal/cpython-unwind

• Issue: Supporting stack unwinding in the JIT compiler #126910

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @diegorusso for commit 5d53a15 🤖

Results will be shown at:

https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F144137%2Fmerge

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

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @diegorusso for commit c179848 🤖

Results will be shown at:

https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F144137%2Fmerge

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

[pablogsal]

This is not used anywhere

[diegorusso]

that's a leftover of a previous logic to test the existence of JIT frames. Will be removing it.

[diegorusso]

I guess you referred only to jit_code_ranges = len(ranges). The rest is needed.

[Fidget-Spinner]

There's https://docs.python.org/3/library/sys.html#sys._jit which you can use instead of this.

[diegorusso]

Unfortunately that API doesn't differentiate between --enable-eperimental-jit and --enable-experimental-jit=interpreter. In both cases I have:

>>> sys._jit.is_enabled()
True
>>> sys._jit.is_available()
True

This is not useful because it doesn't tell me anything about the backend.

For testing purposes I need to know if the backend is just the interpreter because it means that I don't have any memory allocated for the JIT. This is really a corner case but in our CI we do build --enable-experimental-jit=interpreter as a gatekeeper for the JIT jobs.

[pablogsal]

if the module lookup succeeds but filename retrieval fails (len == 0 or len >= MAX_PATH), it falls through without returning anything, This would then check JIT, which is probably fine, but the logic is unclear. Perhaps we can add an explicit return "other"; inside the outer if block after the inner if.

[pablogsal]

Do you get a warning here for interp being unused? I would have expected the compiler to complain for platforms without HAVE_DLADDR or WASI 🤔

[pablogsal]

Super nit (feel free to ignore): extract a global with a name for 0x1000

[pablogsal]

We can see how CI behaves here but when frame pointers aren't expected, asserting exactly 1 Python frame seems brittle. If the unwinder happens to traverse one valid frame before hitting invalid data, this works, but this behavior isn't guaranteed. Consider whether <= 1 perhaps?

[pablogsal]

LGTM! I just left a bunch of minor nits. Feel free to land without those

[pablogsal]

Very cool and great work @diegorusso 💯