gh-100239: Handle NaN and zero division in guards

[Eclips4]

• Issue: Specialize long tail of binary operations using a table. #100239

[iritkatriel]

Shall we check for NaN as well?

[Eclips4]

Shall we check for NaN as well?

You mean "Shall we continue with specialization if we encounter a NaN value?"?
If so, I think we should add a check for NaN to all guards because I don't see any benefit in specializing to get a NaN result (As far as I remember, any operation involving NaN results in NaN).

[iritkatriel]

Shall we check for NaN as well?

You mean "Shall we continue with specialization if we encounter a NaN value?"? If so, I think we should add a check for NaN to all guards because I don't see any benefit in specializing to get a NaN result (As far as I remember, any operation involving NaN results in NaN).

Yes, I guess it would be fine. Might be good to cover this in the tests though.

[iritkatriel]

Doesn't need to be in this PR, but I don't think we have a test for deoptimizing BINARY_OP_EXTEND. We should.

[Eclips4]

Unfortunately ./python.exe -m test -v test_opcache -m test_binary_op -R 3:3 is failing:

Details

==CPython3.14.0a4+ (heads/fix-zero-division-dirty:850a862a7d8, Jan182025, 13:17:14) [Clang16.0.0 (clang-1600.0.26.4)]
==macOS-15.1.1-arm64-arm-64bit-Mach-Olittle-endian==Pythonbuild: debug==cwd:/Users/eclips4/programming/programming-languages/cpython/build/test_python_worker_4826æ==CPUcount: 8==encodings: locale=UTF-8FS=utf-8==resources: alltestresourcesaredisabled, use-uoptiontounskiptestsUsingrandomseed: 5489788770:00:00loadavg: 2.19Run1testsequentiallyinasingleprocess0:00:00loadavg: 2.19 [1/1] test_opcachebeginning6repetitions. Showingnumberofleaks (. for0orless, Xfor10ormore)
123:456test_binary_op (test.test_opcache.TestSpecializer.test_binary_op) ... ok----------------------------------------------------------------------Ran1testin0.006sOKXtest_binary_op (test.test_opcache.TestSpecializer.test_binary_op) ... ok----------------------------------------------------------------------Ran1testin0.007sOKXtest_binary_op (test.test_opcache.TestSpecializer.test_binary_op) ... FAIL======================================================================FAIL: test_binary_op (test.test_opcache.TestSpecializer.test_binary_op)
----------------------------------------------------------------------Traceback (mostrecentcalllast):
File"/Users/eclips4/programming/programming-languages/cpython/Lib/test/test_opcache.py", line1405, intest_binary_opbinary_op_nan()
~~~~~~~~~~~~~^^File"/Users/eclips4/programming/programming-languages/cpython/Lib/test/test_opcache.py", line1402, inbinary_op_nanself.assert_no_opcode(compactlong_lhs, "BINARY_OP_EXTEND")
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"/Users/eclips4/programming/programming-languages/cpython/Lib/test/test_opcache.py", line42, inassert_no_opcodeself.assertNotIn(opname, opnames)
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^AssertionError: 'BINARY_OP_EXTEND'unexpectedlyfoundin {'RESUME_CHECK', 'RETURN_VALUE', 'POP_TOP', 'LOAD_CONST_IMMORTAL', 'BINARY_OP_EXTEND', 'LOAD_SMALL_INT', 'LOAD_FAST'}
----------------------------------------------------------------------Ran1testin0.012sFAILED (failures=1)
testtest_opcachefailedtest_opcachefailed (1failure)
==Testsresult: FAILURE==1testfailed:
test_opcacheTotalduration: 174msTotaltests: run=1 (filtered) failures=1Totaltestfiles: run=1/1 (filtered) failed=1Result: FAILURE

I'm digging into what's happening.

[Eclips4]

Okay, It seems that to get de-specialization it need to "hit" more guards. I'm not sure, though, if this is right.

[iritkatriel]

Thank you!

[Eclips4]

Thank you Irit and Tomas for your review!