gh-92228: disable the compiler's 'small exit block inlining' optimization for blocks that have a line number

[iritkatriel]

Inlining of code that corresponds to source code lines, can make it hard to distinguish later between code which is only reachable from except handlers, and that which is reachable in normal control flow. This caused problems with the debugger's jump feature.

This PR turns off the inlining optimisation for code which has line numbers. We still inline things like the implicit "return None".

• Issue: Cannot perform frame jump after handled exception #92228

Automerge-Triggered-By: GH:tiran

[tiran]

Does the PR need yet another bytecode magic bump or can we piggy-back on the magic bump that was merged yesterday?

[iritkatriel]

Does the PR need yet another bytecode magic bump or can we piggy-back on the magic bump that was merged yesterday?

I think it's safe to piggyback. In any case, we didn't change the set of opcodes so I think the old code would normally still work (unless you ask the compiler to jump over an except block to a small exit block).

[iritkatriel]

Benchmark results show no difference overall:

+-------------------------+----------+------------------------+
| Benchmark               | baseline | gh-92228               |
+=========================+==========+========================+
| chameleon               | 6.37 ms  | 6.27 ms: 1.02x faster  |
+-------------------------+----------+------------------------+
| chaos                   | 64.9 ms  | 65.2 ms: 1.00x slower  |
+-------------------------+----------+------------------------+
| crypto_pyaes            | 72.0 ms  | 74.1 ms: 1.03x slower  |
+-------------------------+----------+------------------------+
| deltablue               | 3.23 ms  | 3.29 ms: 1.02x slower  |
+-------------------------+----------+------------------------+
| dulwich_log             | 60.9 ms  | 61.7 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| float                   | 70.0 ms  | 71.7 ms: 1.02x slower  |
+-------------------------+----------+------------------------+
| go                      | 135 ms   | 133 ms: 1.01x faster   |
+-------------------------+----------+------------------------+
| hexiom                  | 5.95 ms  | 5.98 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| json                    | 4.77 ms  | 4.72 ms: 1.01x faster  |
+-------------------------+----------+------------------------+
| json_dumps              | 12.0 ms  | 12.1 ms: 1.00x slower  |
+-------------------------+----------+------------------------+
| json_loads              | 24.4 us  | 24.5 us: 1.00x slower  |
+-------------------------+----------+------------------------+
| logging_format          | 6.29 us  | 6.34 us: 1.01x slower  |
+-------------------------+----------+------------------------+
| logging_silent          | 89.4 ns  | 90.6 ns: 1.01x slower  |
+-------------------------+----------+------------------------+
| logging_simple          | 5.68 us  | 5.77 us: 1.02x slower  |
+-------------------------+----------+------------------------+
| meteor_contest          | 100 ms   | 101 ms: 1.01x slower   |
+-------------------------+----------+------------------------+
| nbody                   | 92.6 ms  | 90.9 ms: 1.02x faster  |
+-------------------------+----------+------------------------+
| pathlib                 | 17.8 ms  | 18.1 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| pickle                  | 10.3 us  | 10.4 us: 1.00x slower  |
+-------------------------+----------+------------------------+
| pickle_dict             | 31.1 us  | 31.3 us: 1.01x slower  |
+-------------------------+----------+------------------------+
| pickle_list             | 4.30 us  | 4.13 us: 1.04x faster  |
+-------------------------+----------+------------------------+
| pickle_pure_python      | 285 us   | 289 us: 1.01x slower   |
+-------------------------+----------+------------------------+
| pidigits                | 199 ms   | 193 ms: 1.03x faster   |
+-------------------------+----------+------------------------+
| pycparser               | 1.07 sec | 1.11 sec: 1.03x slower |
+-------------------------+----------+------------------------+
| pyflate                 | 398 ms   | 389 ms: 1.02x faster   |
+-------------------------+----------+------------------------+
| python_startup          | 8.20 ms  | 8.26 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| python_startup_no_site  | 6.05 ms  | 6.10 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| raytrace                | 288 ms   | 281 ms: 1.02x faster   |
+-------------------------+----------+------------------------+
| regex_dna               | 200 ms   | 203 ms: 1.01x slower   |
+-------------------------+----------+------------------------+
| regex_effbot            | 3.41 ms  | 3.70 ms: 1.08x slower  |
+-------------------------+----------+------------------------+
| regex_v8                | 21.9 ms  | 21.8 ms: 1.01x faster  |
+-------------------------+----------+------------------------+
| scimark_fft             | 306 ms   | 302 ms: 1.02x faster   |
+-------------------------+----------+------------------------+
| scimark_sor             | 111 ms   | 109 ms: 1.02x faster   |
+-------------------------+----------+------------------------+
| scimark_sparse_mat_mult | 4.00 ms  | 3.90 ms: 1.03x faster  |
+-------------------------+----------+------------------------+
| spectral_norm           | 94.1 ms  | 91.3 ms: 1.03x faster  |
+-------------------------+----------+------------------------+
| sympy_integrate         | 20.1 ms  | 20.1 ms: 1.00x slower  |
+-------------------------+----------+------------------------+
| sympy_sum               | 160 ms   | 159 ms: 1.01x faster   |
+-------------------------+----------+------------------------+
| sympy_str               | 279 ms   | 276 ms: 1.01x faster   |
+-------------------------+----------+------------------------+
| telco                   | 6.55 ms  | 6.44 ms: 1.02x faster  |
+-------------------------+----------+------------------------+
| thrift                  | 737 us   | 727 us: 1.01x faster   |
+-------------------------+----------+------------------------+
| unpack_sequence         | 44.2 ns  | 44.8 ns: 1.01x slower  |
+-------------------------+----------+------------------------+
| unpickle                | 13.2 us  | 13.6 us: 1.03x slower  |
+-------------------------+----------+------------------------+
| unpickle_list           | 5.02 us  | 4.97 us: 1.01x faster  |
+-------------------------+----------+------------------------+
| unpickle_pure_python    | 198 us   | 200 us: 1.01x slower   |
+-------------------------+----------+------------------------+
| xml_etree_iterparse     | 106 ms   | 106 ms: 1.01x slower   |
+-------------------------+----------+------------------------+
| xml_etree_process       | 52.3 ms  | 52.7 ms: 1.01x slower  |
+-------------------------+----------+------------------------+
| Geometric mean          | (ref)    | 1.00x slower           |
+-------------------------+----------+------------------------+

[tiran]

Tests are passing, benchmark shows no impact, C code changes look fine, test changes are plausible. I trust you know what you are doing, too. 👍

[miss-islington]

Thanks @iritkatriel for the PR 🌮🎉.. I'm working now to backport this PR to: 3.11.
🐍🍒⛏🤖

[miss-islington]

Sorry, @iritkatriel, I could not cleanly backport this to 3.11 due to a conflict.
Please backport using cherry_picker on command line.
cherry_picker bde06e1b8381f140b296a397ddd1deb1c784ff8e 3.11

[bedevere-bot]

GH-94643 is a backport of this pull request to the 3.11 branch.

[esc]

This PR turns off the inlining optimisation for code which has line numbers. We still inline things like the implicit "return None".

Quick question: what other "things" could possibly not have a line number besides the implicit return None?