nyash-codex
fa8727c2f8
docs: Phase 196 Select expansion bug fix instruction document
Phase 196 investigation & fix plan for JoinIR→MIR Select conversion bug:
- Goal: Fix "correct PHI in JoinIR → undefined PHI inputs in MIR" issue
- Scope: Bridge layer only (join_ir_vm_bridge / merge line)
Task breakdown:
- 196-1: Minimal reproduction case (phase195_sum_count.hako)
- Record expected JoinIR structure (ValueIds, block IDs)
- Record broken MIR structure (undefined PHI inputs)
- Identify which Select instruction causes the bug
- 196-2: Select expansion path analysis (read-only)
- Locate handle_select in joinir_block.rs
- Understand "1 Select = 3 blocks + 1 PHI" conversion
- Document responsibility locations
- 196-3: Fix strategy (no new boxes)
- Candidate 1: PHI inputs ValueId check (use emit results)
- Candidate 2: Block reuse vs overwrite (Phase 33-20 lesson)
- Candidate 3: JoinInlineBoundary remapper usage (remap before PHI)
- One-at-a-time fix approach
- 196-4: E2E re-verification + regression check
- phase195_sum_count.hako → 72
- loop_if_phi.hako (single-carrier P3)
- P1/P2/P4 representative tests
- 196-5: Documentation updates
- phase196-select-bug-analysis.md (Before/After)
- CURRENT_TASK.md (Phase 196 completion)
- overview (Select expansion line verified)
Design principles:
- Minimal changes (existing Select expansion logic only)
- Focus on one candidate at a time (Fail-Fast)
- Explicit errors (remapper.get_value() must not silently fail)
- Document-driven (record Before/After for future reference)
Success criteria:
- phase195_sum_count.hako outputs 72
- No SSA-undef / PHI errors
- No regressions in P1/P2/P3/P4 tests
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2025-12-09 14:07:42 +09:00
..
2025-09-17 10:58:12 +09:00
2025-12-09 14:07:42 +09:00
2025-11-06 15:41:52 +09:00
2025-11-06 15:41:52 +09:00
2025-08-26 01:42:18 +09:00
2025-09-05 13:29:17 +09:00
2025-12-04 14:19:48 +09:00
2025-12-04 14:19:48 +09:00