refactor(llvm): Phase 131-12-P2 - block_end_values SSOT 化(WIP)

## 実装内容
- get_end_values() API 追加
- _value_at_end_i64() を snapshot-only に変更
- def_blocks 即時更新
- PHI incoming を snapshot から取得

## 発見された問題
- 同一ブロック内の def→use が predecessor snapshot を見てしまう
- これは次フェーズで resolve_cur / resolve_incoming 分離で修正

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
nyash-codex
2025-12-14 21:55:44 +09:00
parent 7dfd6ff1d9
commit eb70dfc5bb
4 changed files with 75 additions and 67 deletions

View File

@ -245,6 +245,12 @@ def lower_blocks(builder, func: ir.Function, block_by_id: Dict[int, Dict[str, An
vmap_cur[dst] = _gval vmap_cur[dst] = _gval
if dst not in created_ids and dst in vmap_cur: if dst not in created_ids and dst in vmap_cur:
created_ids.append(dst) created_ids.append(dst)
# P0-1.5: Update def_blocks IMMEDIATELY after instruction lowering
# This ensures resolver can detect defined_here for same-block uses
try:
builder.def_blocks.setdefault(dst, set()).add(block_data.get("id", 0))
except Exception:
pass
except Exception: except Exception:
pass pass
# Phase 131-4 Pass A: DEFER terminators until after PHI finalization # Phase 131-4 Pass A: DEFER terminators until after PHI finalization

View File

@ -58,39 +58,42 @@ def lower_phi(
# Fallback: use blocks in incoming list # Fallback: use blocks in incoming list
actual_preds = [b for _, b in incoming] actual_preds = [b for _, b in incoming]
# Collect incoming values # P1: Collect incoming values from corresponding snapshots (SSOT)
incoming_pairs: List[Tuple[ir.Block, ir.Value]] = [] incoming_pairs: List[Tuple[ir.Block, ir.Value]] = []
used_default_zero = False used_default_zero = False
import os
strict_mode = os.environ.get('NYASH_LLVM_STRICT') == '1'
for block_id in actual_preds: for block_id in actual_preds:
block = bb_map.get(block_id) block = bb_map.get(block_id)
vid = incoming_map.get(block_id) vid = incoming_map.get(block_id)
if block is None: if block is None:
continue continue
# Prefer resolver-driven localization per predecessor block to satisfy dominance
if vid is not None and resolver is not None and bb_map is not None: # P1: Get value from pred_bid's snapshot (SSOT - no global search)
try: if block_end_values is not None:
pred_block_obj = bb_map.get(block_id) pred_snapshot = block_end_values.get(block_id, {})
if pred_block_obj is not None and hasattr(resolver, 'resolve_i64'): val = pred_snapshot.get(vid) if vid is not None else None
val = resolver.resolve_i64(vid, pred_block_obj, preds_map or {}, block_end_values or {}, vmap, bb_map) else:
else: val = None
val = None
except Exception: if val is None:
val = None # P1: STRICT mode - fail fast on snapshot miss
if val is None: if strict_mode:
# Missing incoming for this predecessor → default 0 available_keys = sorted(list(pred_snapshot.keys())) if block_end_values is not None else []
val = ir.Constant(phi_type, 0) raise RuntimeError(
used_default_zero = True f"[LLVM_PY/STRICT] PHI incoming miss:\n"
f" Source ValueId: v{vid}\n"
f" Predecessor Block: bb{block_id}\n"
f" PHI destination: v{dst_vid}\n"
f" Available in snapshot: {available_keys}\n"
f" Hint: Value v{vid} should be in block_end_values[{block_id}]"
)
# Non-STRICT: fallback to 0
val = ir.Constant(phi_type, 0)
used_default_zero = True
else: else:
# Snapshot fallback
if block_end_values is not None:
snap = block_end_values.get(block_id, {})
val = snap.get(vid) if vid is not None else None
else:
val = vmap.get(vid) if vid is not None else None
if not val:
# Missing incoming for this predecessor → default 0
val = ir.Constant(phi_type, 0)
used_default_zero = True
# Coerce pointer to i64 at predecessor end # Coerce pointer to i64 at predecessor end
if hasattr(val, 'type') and val.type != phi_type: if hasattr(val, 'type') and val.type != phi_type:
pb = ir.IRBuilder(block) pb = ir.IRBuilder(block)

View File

@ -119,6 +119,9 @@ class NyashLLVMBuilder:
# Resolver for unified value resolution # Resolver for unified value resolution
self.resolver = Resolver(self.vmap, self.bb_map) self.resolver = Resolver(self.vmap, self.bb_map)
# P0-1: Connect builder's SSOT structures to resolver
self.resolver.def_blocks = self.def_blocks
self.resolver.block_end_values = self.block_end_values
# Statistics # Statistics
self.loop_count = 0 self.loop_count = 0

View File

@ -61,6 +61,8 @@ class Resolver:
self.def_blocks = {} self.def_blocks = {}
# Optional: block -> { dst_vid -> [(pred_bid, val_vid), ...] } for PHIs from MIR JSON # Optional: block -> { dst_vid -> [(pred_bid, val_vid), ...] } for PHIs from MIR JSON
self.block_phi_incomings = {} self.block_phi_incomings = {}
# P0-1: SSOT for end-of-block values (snapshots)
self.block_end_values = {}
def mark_string(self, value_id: int) -> None: def mark_string(self, value_id: int) -> None:
try: try:
@ -322,28 +324,34 @@ class Resolver:
self.ptr_cache[cache_key] = result self.ptr_cache[cache_key] = result
return result return result
def get_end_values(self, block_id: int) -> Dict[int, ir.Value]:
"""P0-1: Get end-of-block snapshot (SSOT).
Returns the snapshot of values at the end of the specified block.
This is the single source of truth for PHI incoming resolution.
"""
return self.block_end_values.get(block_id, {})
def _value_at_end_i64(self, value_id: int, block_id: int, preds: Dict[int, list], def _value_at_end_i64(self, value_id: int, block_id: int, preds: Dict[int, list],
block_end_values: Dict[int, Dict[int, Any]], vmap: Dict[int, Any], block_end_values: Dict[int, Dict[int, Any]], vmap: Dict[int, Any],
bb_map: Optional[Dict[int, ir.Block]] = None, bb_map: Optional[Dict[int, ir.Block]] = None,
_vis: Optional[set] = None) -> ir.Value: _vis: Optional[set] = None) -> ir.Value:
"""Resolve value as i64 at the end of a given block by traversing predecessors if needed.""" """P0-2: Resolve value from block snapshot only (no global search).
This function now ONLY looks at the block_end_values snapshot for the
specified block. It does NOT recursively search predecessors. This
eliminates processing-order dependency bugs.
"""
trace_phi(f"[resolve] end_i64 enter: bb{block_id} v{value_id}") trace_phi(f"[resolve] end_i64 enter: bb{block_id} v{value_id}")
key = (block_id, value_id) key = (block_id, value_id)
if key in self._end_i64_cache: if key in self._end_i64_cache:
return self._end_i64_cache[key] return self._end_i64_cache[key]
if _vis is None:
_vis = set()
if key in _vis:
trace_phi(f"[resolve] cycle detected at end_i64(bb{block_id}, v{value_id}) → 0")
return ir.Constant(self.i64, 0)
_vis.add(key)
# Do not synthesize PHIs here. Placeholders are created in the function prepass. # P0-2: ONLY use snapshot - no predecessor recursion
# If present in snapshot, coerce there
snap = block_end_values.get(block_id, {}) snap = block_end_values.get(block_id, {})
if value_id in snap and snap[value_id] is not None: val = snap.get(value_id)
val = snap[value_id]
if val is not None:
is_phi_val = False is_phi_val = False
try: try:
is_phi_val = hasattr(val, 'add_incoming') is_phi_val = hasattr(val, 'add_incoming')
@ -361,28 +369,16 @@ class Resolver:
except Exception: except Exception:
belongs_here = False belongs_here = False
if belongs_here: if belongs_here:
self._end_i64_cache[key] = val coerced = self._coerce_in_block_to_i64(val, block_id, bb_map)
return val self._end_i64_cache[key] = coerced
# Otherwise ignore and try predecessors to avoid self-carry from foreign PHI return coerced
coerced = self._coerce_in_block_to_i64(val, block_id, bb_map) # Otherwise, PHI from wrong block -> treat as miss
self._end_i64_cache[key] = coerced else:
return coerced coerced = self._coerce_in_block_to_i64(val, block_id, bb_map)
self._end_i64_cache[key] = coerced
return coerced
# Try recursively from predecessors # P0-2: Snapshot miss - STRICT mode error, else fallback to 0
pred_ids = [p for p in preds.get(block_id, []) if p != block_id]
for p in pred_ids:
v = self._value_at_end_i64(value_id, p, preds, block_end_values, vmap, bb_map, _vis)
if v is not None:
self._end_i64_cache[key] = v
return v
# Do not use global vmap here; if not materialized by end of this block
# (or its preds), bail out with zero to preserve dominance.
preds_s = ','.join(str(x) for x in pred_ids)
trace_phi(f"[resolve] end_i64 miss: bb{block_id} v{value_id} preds=[{preds_s}] → 0")
# P0-2: STRICT mode - fail fast on resolution miss
import os import os
if os.environ.get('NYASH_LLVM_STRICT') == '1': if os.environ.get('NYASH_LLVM_STRICT') == '1':
# Collect diagnostic information # Collect diagnostic information
@ -393,19 +389,19 @@ class Resolver:
except Exception: except Exception:
pass pass
# Build search path for diagnostics snapshot_keys = sorted(list(snap.keys())) if snap else []
search_path = [block_id] + pred_ids
raise RuntimeError( raise RuntimeError(
f"[LLVM_PY/STRICT] PHI resolution miss:\n" f"[LLVM_PY/STRICT] Value not in block snapshot:\n"
f" ValueId: {value_id}\n" f" ValueId: v{value_id}\n"
f" Target block: {block_id}\n" f" Block: bb{block_id}\n"
f" Predecessors: [{preds_s}]\n" f" Available in snapshot: {snapshot_keys}\n"
f" Search path: {search_path}\n"
f" {def_blocks_info}\n" f" {def_blocks_info}\n"
f" Hint: PHI dst may not be registered in def_blocks, or dominance issue" f" Hint: Value should be in block_end_values[{block_id}] but is missing"
) )
# Non-STRICT: fallback to 0 (for diagnostics)
trace_phi(f"[resolve] end_i64 miss: bb{block_id} v{value_id} → 0")
z = ir.Constant(self.i64, 0) z = ir.Constant(self.i64, 0)
self._end_i64_cache[key] = z self._end_i64_cache[key] = z
return z return z