Summary

LOAD_DEREF and LOAD_CLOSURE were missing HAS_FREE_FLAG in the generated opcode metadata, and both also carried HAS_LOCAL_FLAG. As a result dis.hasfree no longer reported them and dis.haslocal wrongly included them — breaking the documented dis.hasfree/dis.haslocal contract. LOAD_CLOSURE regressed in 3.13 (when it became a pseudo-op inheriting LOAD_FAST's flags); LOAD_DEREF regressed in 3.14 (when it began using _PyCell_GetStackRef, which the cases generator's has_free heuristic didn't recognize). This restores the named 3.12 behavior. Surface: introspection metadata only — the generated interpreter/optimizer code is unchanged.

The cases generator now treats _PyCell_GetStackRef as a free-variable access, and the LOAD_CLOSURE pseudo gets an explicit HAS_FREE. A pseudo that explicitly accesses a free variable now suppresses the inherited HAS_LOCAL, mirroring the existing uses_locals and not has_free rule for real instructions (the "HAS_FREE" in self.flags guard is deliberate — the explicit pseudo flags are stored as name strings).

Verification

• make regen-all → git status clean (no generated churn beyond the intended opcode/uop metadata flag lines; generated_cases.c.h, executor_cases.c.h, _opcode_metadata.py unchanged).
• ./python.exe -m test test__opcode test_dis test_compile → SUCCESS (342 tests).
• _PyCell_GetStackRef has exactly one consumer (LOAD_DEREF), so no other opcode's flags change.

Real behavior proof

• Behavior addressed: dis.hasfree / dis.haslocal misclassify LOAD_DEREF and LOAD_CLOSURE.
• Real environment tested: locally-built ./python.exe (--with-pydebug) from this branch on macOS arm64.
• Exact steps or command run after this patch:

  ./python.exe -c "import dis; o=dis.opmap; \
    print('LOAD_DEREF', o['LOAD_DEREF'] in dis.hasfree, o['LOAD_DEREF'] in dis.haslocal); \
    print('LOAD_CLOSURE', o['LOAD_CLOSURE'] in dis.hasfree, o['LOAD_CLOSURE'] in dis.haslocal); \
    print(sorted(dis.opname[x] for x in dis.hasfree))"
  

• Evidence after fix:

  LOAD_DEREF True False
  LOAD_CLOSURE True False
  ['DELETE_DEREF', 'LOAD_CLOSURE', 'LOAD_DEREF', 'LOAD_FROM_DICT_OR_DEREF', 'MAKE_CELL', 'STORE_DEREF']
  

• Observed result after fix: both ops are in dis.hasfree and neither is in dis.haslocal; the hasfree set now lists the six named free/cell opcodes, restoring the named 3.12 classification.
• What was not tested: no runtime/interpreter behavior change exists to test (the changed HAS_FREE/HAS_LOCAL flags feed only _opcode/dis); not built on Windows/Linux or on free-threading/JIT configs (CI will cover those). Backport split noted below.

Backport

The broken-version set differs per opcode: LOAD_DEREF regressed after 3.13; LOAD_CLOSURE is broken on 3.13 too. Requesting needs backport to 3.13 and 3.14.

Note that this diff will not cherry-pick cleanly to 3.13: on 3.13 LOAD_DEREF still uses PyCell_GetRef (already correctly detected as has_free), so the _PyCell_GetStackRef line is inapplicable there. 3.14 takes the full diff; 3.13 likely needs a LOAD_CLOSURE-only manual backport (the pseudo flag + the analyzer.py pseudo change).

• Issue: LOAD_DEREF and LOAD_CLOSURE have incorrect opcode metadata flags #151321