diff --git a/lib/Projectrans/Phase2Translator.py b/lib/Projectrans/Phase2Translator.py index 65976e0..f9040f4 100644 --- a/lib/Projectrans/Phase2Translator.py +++ b/lib/Projectrans/Phase2Translator.py @@ -11,13 +11,14 @@ import json import pickle import time import cProfile +from collections import OrderedDict from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed, wait, FIRST_COMPLETED from lib.core.VLogger import get_logger as _vlog from lib.constants.config import mode as _config_mode from lib.core.Handles.HandlesBase import CTypeInfo from lib.core.SymbolData import ModuleSymbols from lib.includes import t -from lib.core.SymbolUtils import AnnotationContainsName, ExpandSoftKeywords +from lib.core.SymbolUtils import AnnotationContainsName, ExpandSoftKeywords from lib.core.SymbolUtils import ClassHasFuncLevelCVTable, IsVtableSlotMethod from lib.Projectrans.Utils import compute_sha1, _check_annotation_for_export, find_reachable_files_from_entries, topological_sort_files, parse_python_file from lib.Projectrans.Phase1Generator import _TryGlobalCache, _SaveToGlobalCache, _GLOBAL_CACHE_DIR @@ -47,6 +48,65 @@ def _CachedIsFile(path: str) -> bool: return r +# --------------------------------------------------------------- +# 文本读缓存（性能）：同一个文件在一次构建里被**多个阶段反复读** +# （prof 基线 PROF/20260930_224857：_io.open 12,740 次 / self 28.8s、 +# read 11,902 次 / self 30.4s、TextIOWrapper.__exit__ 16.6s —— 128 个模块 +# 却产生上万次读，主要是同一份 .stub.ll / .ll / deps.json 被 +# _compile_one / _recombine_ll / 依赖解析等重复打开）。 +# 键 = (路径, mtime_ns, 大小)：文件被写过或替换后**自动失配**，因此无需手工失效 +# （比"路径→文本"的裸缓存安全）。只缓存 newline=None 的默认换行翻译读， +# 与直接 open(...).read() 语义逐字相同。 +# 内存：按总字节预算 + LRU 淘汰；单文件超过 _TEXT_CACHE_FILE_MAX 不入缓存 +# （合并后的 .ll 可达数十 MB，避免把内存吃光）。 +# --------------------------------------------------------------- +_TEXT_CACHE_MAX_BYTES: int = 64 * 1024 * 1024 +_TEXT_CACHE_FILE_MAX: int = 32 * 1024 * 1024 +_TextReadCache: OrderedDict[tuple[str, int, int], str] = OrderedDict() +_TextReadCacheBytes: int = 0 + + +def _ReadTextCached(path: str, newline: str | None = None) -> str: + """读文本文件（带 LRU 缓存）；newline 非 None 或 stat 失败时退化为直接读。""" + global _TextReadCacheBytes + if newline is None: + try: + _st = os.stat(path) + except OSError: + _st = None + if _st is not None and _st.st_size <= _TEXT_CACHE_FILE_MAX: + _key = (path, _st.st_mtime_ns, _st.st_size) + _hit = _TextReadCache.get(_key) + if _hit is not None: + _TextReadCache.move_to_end(_key) + return _hit + with open(path, 'r', encoding='utf-8') as _f: + _text = _f.read() + _TextReadCache[_key] = _text + _TextReadCacheBytes += len(_text) + while _TextReadCacheBytes > _TEXT_CACHE_MAX_BYTES and len(_TextReadCache) > 1: + _, _old = _TextReadCache.popitem(last=False) + _TextReadCacheBytes -= len(_old) + return _text + with open(path, 'r', encoding='utf-8', newline=newline) as _f: + return _f.read() + + +def _InvalidateTextCache(path: str) -> None: + """丢弃 path 的文本读缓存条目：本进程**改写**该文件后必须调用。 + + 只靠 (mtime_ns, size) 判新旧在 Windows 上不安全：文件时间是系统时钟派生， + 粒度约 15.6ms，「先读 → 同一 tick 内原地改写 → 再读」且改写前后字节数相同 + （实例：_compile_one 前的 type 定义重排，行集合不变仅次序变，长度逐字节相同） + 时键完全一致，会命中过期文本。故写侧必须显式失效，不做时间戳兜底。 + """ + global _TextReadCacheBytes + if not _TextReadCache: + return + for _k in [k for k in _TextReadCache if k[0] == path]: + _TextReadCacheBytes -= len(_TextReadCache.pop(_k)) + + def _FindAsanRuntimeDir() -> str | None: """定位 clang64 的 ASan 运行库目录（按 LLVM 大版本取最新的那个） @@ -153,28 +213,28 @@ _DWARF_TAG_ANY_RE = re.compile(r'DW_TAG_\w+') _DWARF_TAG_SUBPROGRAM_RE = re.compile(r'DW_TAG_subprogram\b') _DWARF_LOWPC_RE = re.compile(r'DW_AT_low_pc\s*\(\s*(0x[0-9A-Fa-f]+)\s*\)') _DWARF_HIGHPC_RE = re.compile(r'DW_AT_high_pc\s*\(\s*(0x[0-9A-Fa-f]+)\s*\)') -_DWARF_ATNAME_RE = re.compile(r'DW_AT_name\s*\("((?:[^"\\]|\\.)*)"\)') - -# --------------------------------------------------------------- -# `_split_ll` 的逐行正则**预编译**（性能：该函数对合并后的 .ll 逐行跑两遍， -# prof 实测 43.9万行级输入下 re 模块级函数的缓存查找（re._compile）达 1970 万次 -# ≈ 7s，另有 3100 万次 re.Pattern.match；每次 `re.match(字面量, ...)` 都要做 -# "模式串哈希 → 缓存查找"。预编译成模块级常量后省掉这层查找，且匹配路径更短。 -# 语义与逐字写法完全一致（模式串一字未改）。 -# --------------------------------------------------------------- -_RE_DEFINE_SHA1: re.Pattern[str] = re.compile(r'define\s+[^@]*@"?([a-f0-9]+)\.') -_RE_DEFINE_NAME: re.Pattern[str] = re.compile(r'define\s+[^@]*@"?([^"\s(]+)"?') -_RE_DECLARE_NAME: re.Pattern[str] = re.compile(r'declare\s+[^@]*@"?([^"\s(]+)"?') -_RE_LLVM_INTRINSIC: re.Pattern[str] = re.compile(r'@("?)llvm\.') -_RE_AT_NAME: re.Pattern[str] = re.compile(r'(@\S+)') -_RE_SHA1_16_DOT: re.Pattern[str] = re.compile(r'^[a-f0-9]{16}\.') -_RE_DEFINE_PREFIX: re.Pattern[str] = re.compile(r'^define\s+') -_RE_DEFINE_LINKAGE: re.Pattern[str] = re.compile( - r'\b(linkonce_odr|weak_odr|linkonce|weak|common|appending|internal|private)\s+') -_RE_ALWAYSINLINE: re.Pattern[str] = re.compile(r'\balwaysinline\s+') -# `_compile_one` 的逐行符号扫描/全局去重里用到的两条（同一预编译理由） -_RE_AT_EXTERNAL: re.Pattern[str] = re.compile(r'@\S+\s*=\s*external\s+') -_RE_DECLARE_FULL: re.Pattern[str] = re.compile( +_DWARF_ATNAME_RE = re.compile(r'DW_AT_name\s*\("((?:[^"\\]|\\.)*)"\)') + +# --------------------------------------------------------------- +# `_split_ll` 的逐行正则**预编译**（性能：该函数对合并后的 .ll 逐行跑两遍， +# prof 实测 43.9万行级输入下 re 模块级函数的缓存查找（re._compile）达 1970 万次 +# ≈ 7s，另有 3100 万次 re.Pattern.match；每次 `re.match(字面量, ...)` 都要做 +# "模式串哈希 → 缓存查找"。预编译成模块级常量后省掉这层查找，且匹配路径更短。 +# 语义与逐字写法完全一致（模式串一字未改）。 +# --------------------------------------------------------------- +_RE_DEFINE_SHA1: re.Pattern[str] = re.compile(r'define\s+[^@]*@"?([a-f0-9]+)\.') +_RE_DEFINE_NAME: re.Pattern[str] = re.compile(r'define\s+[^@]*@"?([^"\s(]+)"?') +_RE_DECLARE_NAME: re.Pattern[str] = re.compile(r'declare\s+[^@]*@"?([^"\s(]+)"?') +_RE_LLVM_INTRINSIC: re.Pattern[str] = re.compile(r'@("?)llvm\.') +_RE_AT_NAME: re.Pattern[str] = re.compile(r'(@\S+)') +_RE_SHA1_16_DOT: re.Pattern[str] = re.compile(r'^[a-f0-9]{16}\.') +_RE_DEFINE_PREFIX: re.Pattern[str] = re.compile(r'^define\s+') +_RE_DEFINE_LINKAGE: re.Pattern[str] = re.compile( + r'\b(linkonce_odr|weak_odr|linkonce|weak|common|appending|internal|private)\s+') +_RE_ALWAYSINLINE: re.Pattern[str] = re.compile(r'\balwaysinline\s+') +# `_compile_one` 的逐行符号扫描/全局去重里用到的两条（同一预编译理由） +_RE_AT_EXTERNAL: re.Pattern[str] = re.compile(r'@\S+\s*=\s*external\s+') +_RE_DECLARE_FULL: re.Pattern[str] = re.compile( r'^declare\s+(?:dso_local\s+)?[^@]*@\"?([^"\s\(]+)\"?\s*\(') @@ -785,8 +845,7 @@ def _BuildPyiCacheFile(temp_dir: str) -> str | None: sig_path: str = os.path.join(temp_dir, _PYI_CACHE_SIG_NAME) try: if _CachedIsFile(pkl_path) and _CachedIsFile(sig_path): - with open(sig_path, 'r', encoding='utf-8') as f: - stored = json.load(f) + stored = json.loads(_ReadTextCached(sig_path)) # json 往返后 tuple→list，需规范化再比较 if stored == [[n, s, m] for n, s, m in sig]: return pkl_path @@ -814,8 +873,8 @@ def _BuildPyiCacheFile(temp_dir: str) -> str | None: cache[p] = old_entry[2] continue try: - with open(p, 'r', encoding='utf-8') as f: - cache[p] = ast.parse(f.read(), p) + # 走 LRU 文本缓存：同一 .vpi 在增量复用/多 worker 预热路径会被反复读 + cache[p] = ast.parse(_ReadTextCached(p), p) except Exception: cache[p] = None try: @@ -854,8 +913,7 @@ def _PyiCacheWorkerInitializer(temp_dir: str) -> None: new_cache: dict[str, object] = {} for p in _ScanPyiFiles(temp_dir): try: - with open(p, 'r', encoding='utf-8') as f: - new_cache[p] = ast.parse(f.read(), p) + new_cache[p] = ast.parse(_ReadTextCached(p), p) except Exception: new_cache[p] = None if new_cache: @@ -887,18 +945,18 @@ def _count_braces_outside_strings(text: str) -> int: return count -# LLVM IR "无引号标识符"的精确规则（LangRef "Identifiers"；实现即 -# LLLexer::ReadVarName）：首字符 ∈ [-a-zA-Z$._]，后续字符 ∈ [-a-zA-Z$._0-9]。 -# 首字符**不允许是数字**：`$884f3d94ea4d0123.md5_F` 会让 LLLexer::LexDollar -# 走到 ReadVarName() 返回 false ⇒ 产出 error token ⇒ 顶层解析报 -# "expected top-level entity"。故 comdat 名必须以同一规则判定是否加引号 -# （与 LLVM 自身 AsmWriter 的输出形态一致：`$"884f3d94ea4d0123.md5_F"`）。 -# 另：引号名 @"..." 不支持转义，名字含 `"`/`\` 时两种形态都非法 ⇒ 见 -# _AddWeakComdat 内的 fail-fast（不静默替换、不跳过）。 -_LLVM_UNQUOTED_NAME_RE: re.Pattern[str] = re.compile( - r'[-a-zA-Z$._][-a-zA-Z$._0-9]*\Z') - - +# LLVM IR "无引号标识符"的精确规则（LangRef "Identifiers"；实现即 +# LLLexer::ReadVarName）：首字符 ∈ [-a-zA-Z$._]，后续字符 ∈ [-a-zA-Z$._0-9]。 +# 首字符**不允许是数字**：`$884f3d94ea4d0123.md5_F` 会让 LLLexer::LexDollar +# 走到 ReadVarName() 返回 false ⇒ 产出 error token ⇒ 顶层解析报 +# "expected top-level entity"。故 comdat 名必须以同一规则判定是否加引号 +# （与 LLVM 自身 AsmWriter 的输出形态一致：`$"884f3d94ea4d0123.md5_F"`）。 +# 另：引号名 @"..." 不支持转义，名字含 `"`/`\` 时两种形态都非法 ⇒ 见 +# _AddWeakComdat 内的 fail-fast（不静默替换、不跳过）。 +_LLVM_UNQUOTED_NAME_RE: re.Pattern[str] = re.compile( + r'[-a-zA-Z$._][-a-zA-Z$._0-9]*\Z') + + def _AddWeakComdat(ll_content: str) -> str: """为 weak/linkonce 系列 函数/全局 定义补 COMDAT（COFF/PE 必需）。 @@ -1023,17 +1081,17 @@ class Phase2Translator: if os.path.isdir(_clang64_bin): os.environ['PATH'] = _clang64_bin + os.pathsep + os.environ.get('PATH', '') self.asan_rt_dir = _FindAsanRuntimeDir() - self.slice_level = 3 - # 函数级 DWARF 调试元数据（--dwarf；--asan 时自动开启）：False=未启用 - # （零影响，不构造任何 DI 元数据）；True=App 源模块的 .ll 携带 - # !llvm.dbg.cu + 每函数 DISubprogram + 入口 DILocation，供 ASan / - # llvm-symbolizer 反查 ".vp:函数起始行"。 - self.dwarf_enabled: bool = dwarf_enabled - # Ctrl+C 调用栈（--traceback）：False=未启用（零影响，不调用任何外部 - # 工具、不写任何旁文件）；True=链接成功后额外用 llvm-dwarfdump 把最终 - # exe 的 DWARF 行表导出为 ".stackmap"（地址→源行旁文件，供 exe 内 - # 运行时在 Ctrl+C/Ctrl+Break 时逐帧翻译返回地址）。要求 --dwarf 已开启 - # （Config 侧联动，否则行表为空），且目标为 Windows PE。 + self.slice_level = 3 + # 函数级 DWARF 调试元数据（--dwarf；--asan 时自动开启）：False=未启用 + # （零影响，不构造任何 DI 元数据）；True=App 源模块的 .ll 携带 + # !llvm.dbg.cu + 每函数 DISubprogram + 入口 DILocation，供 ASan / + # llvm-symbolizer 反查 ".vp:函数起始行"。 + self.dwarf_enabled: bool = dwarf_enabled + # Ctrl+C 调用栈（--traceback）：False=未启用（零影响，不调用任何外部 + # 工具、不写任何旁文件）；True=链接成功后额外用 llvm-dwarfdump 把最终 + # exe 的 DWARF 行表导出为 ".stackmap"（地址→源行旁文件，供 exe 内 + # 运行时在 Ctrl+C/Ctrl+Break 时逐帧翻译返回地址）。要求 --dwarf 已开启 + # （Config 侧联动，否则行表为空），且目标为 Windows PE。 self.traceback_enabled: bool = traceback_enabled # 调用点热点统计（--hotspot-sites）：None=未启用（零影响）；str=调用点计数 # 落盘文件路径（每函数计时落同目录 hotspot_time.txt）。各模块运行时自带 @@ -1543,8 +1601,7 @@ class Phase2Translator: if exists: py_sha1 = Phase2Translator._include_sha1_cache.get(py_path) if py_sha1 is None: - with open(py_path, 'r', encoding='utf-8') as f: - py_sha1 = compute_sha1(f.read()) + py_sha1 = compute_sha1(_ReadTextCached(py_path)) Phase2Translator._include_sha1_cache[py_path] = py_sha1 actual_sha1 = py_sha1 break @@ -1795,14 +1852,14 @@ class Phase2Translator: deps[mod_name] = dep_sha1 with open(deps_path, 'w', encoding='utf-8') as f: json.dump(deps, f) + _InvalidateTextCache(deps_path) def _LoadDeps(self, sha1: str) -> dict[str, str] | None: """加载依赖指纹""" deps_path = os.path.join(self.output_dir, f"{sha1}.deps.json") if _CachedIsFile(deps_path): try: - with open(deps_path, 'r', encoding='utf-8') as f: - return json.load(f) + return json.loads(_ReadTextCached(deps_path)) except Exception as _e: if _config_mode == "strict": raise @@ -1835,10 +1892,8 @@ class Phase2Translator: if not _CachedIsFile(stub_path) or not _CachedIsFile(text_path): return False - with open(stub_path, 'r', encoding='utf-8') as f: - stub_content = f.read() - with open(text_path, 'r', encoding='utf-8') as f: - text_content = f.read() + stub_content = _ReadTextCached(stub_path) + text_content = _ReadTextCached(text_path) sha1_replacements = {} for mod_name, old_sha1 in saved_deps.items(): @@ -1862,6 +1917,9 @@ class Phase2Translator: f.write(stub_content) with open(text_path, 'w', encoding='utf-8', newline='\n') as f: f.write(text_content) + _InvalidateTextCache(ll_path) + _InvalidateTextCache(stub_path) + _InvalidateTextCache(text_path) self._save_deps(sha1, current_ModuleSha1Map) @@ -1961,8 +2019,7 @@ class Phase2Translator: # 顺序模式与 worker 模式的每文件粒度皆覆盖，见 MarkStubSetDirty）。 from lib.core.Handles.HandlesImports import ImportHandle as _IH _IH.MarkStubSetDirty() - with open(src_path, 'r', encoding='utf-8') as f: - code = f.read() + code = _ReadTextCached(src_path) # SHA1 以主流程 out_path 文件名为准（= 主流程 compute_sha1(parse_python_file 内容)）， # 必须在自动导入注入之前确定。不可在此重读文件重算：主流程收集（need_translate） @@ -2008,7 +2065,7 @@ class Phase2Translator: trans.translator.SliceCount = 0 trans.translator.SliceInfos = [] trans.translator.LlvmGen = None - trans.translator._module_sha1 = sha1 + trans.translator._module_sha1 = sha1 trans.translator._line_offset = _line_offset trans.translator._ModuleSha1Map = ModuleSha1Map trans.translator._struct_sha1_map = self.struct_sha1_map @@ -2054,8 +2111,7 @@ class Phase2Translator: if sha1_key == sha1: continue try: - with open(sig_path, 'r', encoding='utf-8') as f: - sig_content = f.read() + sig_content = _ReadTextCached(sig_path) sig_tree = ast.parse(sig_content, sig_path) for node in ast.iter_child_nodes(sig_tree): if isinstance(node, ast.FunctionDef): @@ -2117,8 +2173,7 @@ class Phase2Translator: continue if os.path.exists(stub_path): try: - with open(stub_path, 'r', encoding='utf-8') as f: - stub_content = f.read() + stub_content = _ReadTextCached(stub_path) for line in stub_content.splitlines(): line = line.strip() if line.startswith('@') and '= global' in line and 'i32' in line and not line.startswith('@llvm'): @@ -2229,16 +2284,19 @@ class Phase2Translator: _vlog().warning(f"后处理添加属性失败: 未找到函数 {func_name} 的定义") with open(out_path, 'w', encoding='utf-8', newline='\n') as f: f.write(result) + _InvalidateTextCache(out_path) stub_content, text_content = self._split_ll(result) base = os.path.splitext(out_path)[0] with open(base + '.stub.ll', 'w', encoding='utf-8', newline='\n') as f: f.write(stub_content) + _InvalidateTextCache(base + '.stub.ll') # 新 stub 已落盘 → 置脏，使本文件后续（及下个文件）的索引 miss 能重扫到它 from lib.core.Handles.HandlesImports import ImportHandle as _IH2 _IH2.MarkStubSetDirty() with open(base + '.text.ll', 'w', encoding='utf-8', newline='\n') as f: f.write(text_content) + _InvalidateTextCache(base + '.text.ll') self._save_deps(sha1, ModuleSha1Map) # 进度日志由调用方统一打印（主进程按文件粒度 [x/total]），此处不再重复输出。 @@ -2248,8 +2306,7 @@ class Phase2Translator: out_dir = os.path.dirname(out_path) if out_dir: os.makedirs(out_dir, exist_ok=True) - with open(stub_path, 'r', encoding='utf-8') as f: - stub_content = f.read() + stub_content = _ReadTextCached(stub_path) with open(out_path, 'w', encoding='utf-8', newline='\n') as f: f.write(stub_content) # 进度日志由调用方统一打印。 @@ -2302,6 +2359,7 @@ class Phase2Translator: if fixed != text: with open(ll_path, 'w', encoding='utf-8', newline='') as f: f.write(fixed) + _InvalidateTextCache(ll_path) return ll_path def _llc_compile(self, ll_path: str, obj_path: str, @@ -2394,8 +2452,7 @@ class Phase2Translator: has_inline = False for ll_path in all_ll_files: try: - with open(ll_path, 'r', encoding='utf-8') as f: - content = f.read() + content = _ReadTextCached(ll_path) if 'alwaysinline' in content: has_inline = True break @@ -2409,8 +2466,7 @@ class Phase2Translator: _vlog().info(" [内联] 检测到 alwaysinline 函数，执行跨模块内联优化...") for ll_path in all_ll_files: try: - with open(ll_path, 'r', encoding='utf-8') as f: - content = f.read() + content = _ReadTextCached(ll_path) type_defs = [] type_def_indices = set() for i, line in enumerate(content.splitlines()): @@ -2432,6 +2488,7 @@ class Phase2Translator: fixed = ''.join(lines2[:header_end]) + '\n'.join(type_defs) + '\n' + ''.join(lines2[header_end:]) with open(ll_path, 'w', encoding='utf-8') as f: f.write(fixed) + _InvalidateTextCache(ll_path) except Exception as e: _vlog().warning(f"Phase2Translator: 忽略异常 {e}", exc_info=e) merged_ll = os.path.join(self.output_dir, '_merged.ll') @@ -2450,8 +2507,7 @@ class Phase2Translator: has_inline = False else: _vlog().info(" [内联] 跨模块内联优化完成") - with open(optimized_ll, 'r', encoding='utf-8') as f: - opt_content = f.read() + opt_content = _ReadTextCached(optimized_ll) opt_content = re.sub(r'^(define|declare)\s+(private|internal|available_externally|linkonce|weak|common|appending|extern_weak|linkonce_odr|weak_odr|external)\s+', r'\1 \2 dso_local ', opt_content, flags=re.MULTILINE) opt_content = re.sub(r'^(define|declare)\s+(?!private|internal|available_externally|linkonce|weak|common|appending|extern_weak|linkonce_odr|weak_odr|external|dso_local|@)', r'\1 dso_local ', opt_content, flags=re.MULTILINE) # LLVM 22+: external 不允许在带初值的全局定义上显式写出，改为 linkonce 保留多模块合并语义 @@ -2490,8 +2546,7 @@ class Phase2Translator: if _CachedIsFile(obj_path): return (rel, 'cached', '') try: - with open(ll_path, 'r', encoding='utf-8') as f: - ll_content = f.read() + ll_content = _ReadTextCached(ll_path) sha1 = os.path.basename(ll_path)[:-3] deps = self._LoadDeps(sha1) stub_prefix = '' @@ -2571,8 +2626,7 @@ class Phase2Translator: if not _CachedIsFile(stub_path): stub_path = os.path.join(self.temp_dir, f"{dep_sha1}.stub.ll") if _CachedIsFile(stub_path): - with open(stub_path, 'r', encoding='utf-8') as sf: - stub_content = sf.read() + stub_content = _ReadTextCached(stub_path) for line in stub_content.splitlines(): stripped = line.strip() if stripped.startswith(';') or stripped.startswith('target ') or stripped.startswith('source_filename') or not stripped: @@ -3369,13 +3423,12 @@ class Phase2Translator: for other_sha1, other_stub in self.stub_files.items(): if os.path.exists(other_stub): try: - with open(other_stub, 'r', encoding='utf-8') as sf: - for sline in sf: - sl = sline.strip() - if sl.startswith('%') and '= type' in sl and 'opaque' not in sl: - struct_name = Phase2Translator._extract_llvm_type_name(sl) - if struct_name: - stub_decls.append(sl) + for sline in _ReadTextCached(other_stub).splitlines(True): + sl = sline.strip() + if sl.startswith('%') and '= type' in sl and 'opaque' not in sl: + struct_name = Phase2Translator._extract_llvm_type_name(sl) + if struct_name: + stub_decls.append(sl) except Exception as e: _vlog().warning(f"Phase2Translator: 忽略异常 {e}", exc_info=e) all_type_defs = {} @@ -3545,8 +3598,7 @@ class Phase2Translator: _vlog().error(f" 翻译未生成 .ll 文件: {module_name}.py") continue - with open(ll_path, 'r', encoding='utf-8') as f: - ll_content = f.read() + ll_content = _ReadTextCached(ll_path) # --traceback 提醒：include 模块若发射了 define 却不带 DI（!llvm.dbg.cu）， # 其 .debug_line 行为空 ⇒ 运行时 Ctrl+C 栈里该 include 帧取不到源行 @@ -3664,6 +3716,7 @@ class Phase2Translator: dso_ll_path = ll_path[:-3] + '_dso.ll' with open(dso_ll_path, 'w', encoding='utf-8') as f: f.write(ll_content) + _InvalidateTextCache(dso_ll_path) compile_tasks.append((dso_ll_path, obj_path, src_path, sha1, module_name, ll_path, precompiled_obj, ll_content)) if not compile_tasks: @@ -3910,12 +3963,12 @@ class Phase2Translator: _TryGlobalCache(sha1, 'pyi', sig_path) if not _CachedIsFile(sig_path): try: - with open(src_path, 'r', encoding='utf-8') as f: - py_content = f.read() + py_content = _ReadTextCached(src_path) sig_content = PythonToStubConverter.convert(py_content, mod_full) os.makedirs(self.temp_dir, exist_ok=True) with open(sig_path, 'w', encoding='utf-8', newline='\n') as f: f.write(sig_content) + _InvalidateTextCache(sig_path) _SaveToGlobalCache(sha1, 'pyi', sig_path) except Exception as e: if _config_mode == "strict": @@ -4137,8 +4190,7 @@ class Phase2Translator: for sha1_key, stub_path in self.stub_files.items(): if os.path.exists(stub_path): try: - with open(stub_path, 'r', encoding='utf-8') as f: - stub_content = f.read() + stub_content = _ReadTextCached(stub_path) for line in stub_content.splitlines(): line = line.strip() if line.startswith('@') and '= global' in line and 'i32' in line and not line.startswith('@llvm'): @@ -4239,8 +4291,7 @@ class Phase2Translator: if sha1_key in shared_doc_meta: continue try: - with open(os.path.join(_GLOBAL_CACHE_DIR, fname), 'r', encoding='utf-8') as f: - data = json.load(f) + data = json.loads(_ReadTextCached(os.path.join(_GLOBAL_CACHE_DIR, fname))) if isinstance(data, dict): shared_doc_meta[sha1_key] = {str(k): str(v) for k, v in data.items()} except Exception as _e: @@ -4621,6 +4672,7 @@ SECTIONS { """ with open(linker_ld_dst, 'w', encoding='utf-8', newline='\n') as f: f.write(default_ld) + _InvalidateTextCache(linker_ld_dst) _vlog().info(f" [链接] 自动创建默认链接脚本: linker.ld") try: